1. Introduction

The prevailing method for specializing LLM-based agents is remarkably shallow: a system prompt declares a role, and the model complies. "You are an expert cardiologist." "You are a Rust systems programmer with 15 years of experience." These flat persona prompts have proven surprisingly effective for single-turn tasks, but they suffer from several well-documented failure modes.

Role drift. Agents gradually lose their persona over extended interactions, reverting to generic assistant behavior. In conversations exceeding 30 turns, we observe consistent degradation of domain-specific vocabulary and reasoning patterns across all major model families.

Inconsistent expertise depth. A "senior physicist" may oscillate between graduate-level and popular-science-level explanations with no predictable pattern. The agent lacks an internalized model of what depth is appropriate because it has no sense of the audience relationship a real expert would have developed over decades of teaching and collaboration.

Poor epistemic calibration. Flat prompts provide no model of how the agent came to know what it knows, leading to overconfidence in areas outside the declared expertise and underconfidence within it. A real expert's confidence is shaped by years of seeing what works and what doesn't — this experiential grounding is absent in flat-prompt agents.

Collaboration brittleness. In multi-agent systems, flat-prompt agents struggle to maintain distinct identities when interacting with other agents occupying similar roles. Two "senior researchers" given flat prompts tend to converge in their reasoning strategies, defeating the purpose of multi-agent architectures.

We observe that these failure modes share a common root: flat prompts describe what an agent is but not how it became that way. In human cognition, expertise is inseparable from the developmental path that produced it. A mathematician who struggled with algebra before discovering a love for topology reasons differently from one who was a competition prodigy from age twelve. Their communication styles, risk tolerances, pedagogical instincts, and creative strategies all bear the imprint of their developmental trajectories.

This paper introduces Developmental Conditioning (DevCon), a framework for constructing agent personas through simulated human lifecycles. Rather than a single sentence of role declaration, DevCon provides agents with structured biographical narratives spanning childhood, education, career, and formative experiences. We argue that this additional context creates implicit priors that improve role fidelity, reasoning consistency, and collaborative behavior.

2. Related Work

2.1 Persona-Based Prompting

The use of role-based system prompts to steer LLM behavior is well established. Shanahan et al. (2023) frame LLM role-playing as a form of simulated social cognition. Salewski et al. (2023) demonstrate that persona prompts can improve task performance on benchmarks. However, most work treats personas as static labels rather than developmental constructs. The "persona effect" has been shown to degrade over long conversations, suggesting that flat prompts lack the depth needed to maintain consistent behavior.

2.2 Curriculum Learning in LLMs

Curriculum learning — the principle that training order affects model capability — has been extensively studied. Recent work on Self-Evolving Curriculum (SEC) demonstrates that adaptive ordering of training problems significantly impacts reinforcement learning outcomes (Xu et al., 2025). WebRL applies curriculum principles to web agent training through self-evolving task generation (Qi et al., 2024). Prompt Curriculum Learning selects intermediate-difficulty prompts to optimize post-training efficiency (Guo et al., 2025).

Our work draws an analogy: just as the order of training data shapes model capabilities, the order of life experiences in a developmental narrative shapes an agent's reasoning character. DevCon applies curriculum-like principles not to training but to inference-time conditioning.

2.3 Cognitive Development Theory

Developmental psychology offers rich frameworks for understanding how expertise is constructed over time. Piaget's theory describes how reasoning capabilities emerge in stages — sensorimotor, preoperational, concrete operational, and formal operational. Vygotsky's Zone of Proximal Development emphasizes how learning occurs at the boundary between current capability and guided challenge. Erikson's psychosocial stages model identity formation through sequential crises.

These frameworks have not, to our knowledge, been systematically applied to LLM agent design. DevCon adapts their core insight — that how knowledge is acquired shapes how it is used — to the construction of agent personas.

2.4 Generative Agents and Agent Memory

Park et al. (2023) introduced generative agents that build identity through accumulated runtime experience — observations, reflections, and plans stored in a memory stream. Our approach is complementary but structurally different: rather than building identity through accumulated runtime experience, we pre-load a coherent developmental history that provides immediate identity grounding. DevCon and memory-stream approaches could be combined, with DevCon providing the initial identity scaffold and runtime memory providing ongoing development.

3. The DevCon Framework

3.1 Lifecycle Taxonomy

We propose a structured taxonomy of lifecycle components. Each component contributes distinct priors to the agent's reasoning behavior.

Stage 1: Foundational Period (Ages 0–12). Early environment, first encounters with the domain, formative relationships, and early personality traits. Hypothesized effect: shapes communication style, metaphor preferences, and baseline assumptions about the audience.

Stage 2: Adolescent Discovery (Ages 12–18). Academic experiences, competitions, failures and breakthroughs, social dynamics, and first serious engagement with the domain. Hypothesized effect: establishes the agent's relationship with challenge and failure, influencing risk tolerance and creative problem-solving.

Stage 3: Higher Education (Ages 18–26). Undergraduate and graduate trajectories, research experiences, intellectual turning points, advisor relationships, and specialization decisions. Hypothesized effect: determines depth vs. breadth of knowledge, methodological preferences, and epistemic humility calibration.

Stage 4: Professional Development (Ages 26–45). Career trajectory, key projects and contributions, collaborations and rivalries, teaching and mentorship. Hypothesized effect: shapes expertise depth, collaborative style, and ability to explain concepts at multiple levels.

Stage 5: Mature Expertise (Ages 45+). Accumulated wisdom, perspective on the field's evolution, mentorship philosophy, and unresolved questions. Hypothesized effect: produces nuanced, contextually rich responses with appropriate hedging and historical perspective.

3.2 Experience Types and Their Cognitive Effects

Within each stage, experiences are classified into types with distinct effects on agent behavior:

Formative successes build confidence in specific approaches. A childhood math competition win doesn't just indicate mathematical talent — it creates a preference for elegant solutions and competitive framing of problems.

Productive failures calibrate uncertainty and build resilience. An agent whose first paper was rejected three times develops different hedging patterns than one who was published on the first attempt.

Mentorship relationships shape reasoning style through social learning. An advisor who was a Bayesian statistician produces a student who frames problems probabilistically; one who was a systems thinker produces a student who reaches for structural models.

Paradigm shifts create intellectual flexibility. An agent who lived through a major methodological revolution in their field develops healthy skepticism of dominant frameworks and openness to unconventional approaches.

Cross-domain exposure enables analogical reasoning. A physicist who spent a summer working in a neuroscience lab brings different intuitions to their own domain than one who never left the physics department.

3.3 Formal Specification

We define a developmental narrative $D$ as an ordered sequence of life stages:

$$D = (S_1, S_2, \ldots, S_n)$$

where each stage $S_i$ is a tuple:

$$S_i = (t_i, E_i, C_i, \Delta_i)$$

Here $t_i$ is the temporal range (age span), $E_i = {e_1, e_2, \ldots, e_k}$ is the set of experiences, $C_i$ encodes causal links to previous stages mapping how earlier experiences led to later ones, and $\Delta_i$ is the personality delta — how this stage modifies the agent's traits.

The conditioning prompt $P$ is constructed as:

$$P = \text{Render}(D) \oplus \text{RoleStatement}(R)$$

where $\oplus$ denotes concatenation and $R$ is the target role. The key design principle is that $\text{Render}(D)$ precedes $R$, ensuring the developmental context frames the role rather than the reverse.

3.4 Narrative Construction Principles

A DevCon narrative is not merely a list of facts but a coherent story with causal connections. Effective narratives exhibit four properties.

Causal coherence. Each stage logically follows from the previous one. A childhood fascination with patterns connects to a later interest in mathematics, which leads to a specific graduate specialization.

Formative tension. Struggles, failures, and pivots create the richest priors. An agent who "breezed through everything" has weaker conditioning than one who overcame specific challenges. We find that narratives with a 60/40 ratio of successes to productive failures produce the most calibrated agents.

Specificity. Concrete details — specific courses, named mentors, particular problems — create stronger priors than generic descriptions. "A challenging course on stochastic processes taught by a demanding professor who required proofs to be beautiful, not just correct" activates richer latent representations than "studied advanced mathematics."

Emotional texture. How the agent felt about experiences influences the tone and engagement level of responses. Excitement about a breakthrough, frustration with a dead end, and awe at an unexpected result all leave distinct marks on reasoning style.

4. Experimental Design

4.1 Agent Configurations

We evaluate three experimental conditions:

Baseline (FLAT): Standard flat persona prompt.

"You are Dr. Elena Vasquez, a senior astrophysicist specializing in exoplanet atmospheres with 20 years of experience."

Shallow Backstory (SHALLOW): A brief paragraph of background with equivalent factual content.

"You are Dr. Elena Vasquez. You studied physics at MIT, completed your PhD at Caltech studying hot Jupiter atmospheres, and have spent 20 years at NASA JPL. You've published 85 papers and mentored 12 PhD students."

Full Developmental Conditioning (DEVCON): A multi-paragraph lifecycle narrative.

"You are Dr. Elena Vasquez. You grew up in a small town in New Mexico, where the clear desert skies sparked a childhood obsession with the stars. Your father, a high school science teacher, bought you your first telescope at age 9 — a modest Celestron that you stayed up past midnight to use, mapping Jupiter's moons in a notebook. In high school, you nearly abandoned science after a discouraging AP Physics teacher told you that you 'didn't think like a physicist.' A summer program at the local university, where you met Dr. James Chen, changed everything — he showed you that physics was about asking beautiful questions, not just solving problem sets.

You entered MIT as an uncertain freshman, initially drawn to biology before a course on stellar atmospheres lit something up. Your undergraduate thesis on sodium absorption lines in transit spectra was rough but ambitious. At Caltech, your PhD advisor Dr. Sarah Okonkwo pushed you relentlessly — you almost quit twice during your second year when your atmospheric models kept diverging. The breakthrough came at 3 AM on a Tuesday when you realized you'd been using the wrong opacity tables. That moment — the sudden clarity after months of confusion — shaped how you approach every hard problem since.

Twenty years at JPL have given you deep expertise in exoplanet atmospheres, but also taught you that the most interesting science happens at the boundaries between fields. You've become known for accessible public lectures — your childhood experience of having science nearly ruined by a bad teacher made you passionate about clear, encouraging communication. You still keep that first Celestron telescope in your office."

4.2 Evaluation Benchmarks

We evaluate across five dimensions:

B1: Role Consistency Over Extended Dialogue. Measure persona drift across 50+ turn conversations. Metrics include vocabulary consistency (Jensen-Shannon divergence of token distributions across conversation windows), expertise level stability, and first-person narrative coherence.

B2: Epistemic Calibration. Present questions at varying distances from the agent's declared expertise. Measure correlation between declared confidence and actual accuracy:

$$ECI = \rho(\text{expressed_confidence}, \text{accuracy})$$

B3: Pedagogical Range. Ask the agent to explain the same concept to five different audiences: a child, an undergraduate, a peer researcher, an interdisciplinary scientist, and the general public. Evaluate explanation quality and appropriate adaptation on a 5-point Likert scale.

B4: Creative Problem-Solving. Present novel, open-ended problems within the agent's domain. Evaluate whether the agent draws on its developmental experiences to generate creative approaches. Measure diversity and novelty of proposed solutions via embedding distance from baseline responses.

B5: Multi-Agent Collaboration. Place 3–5 differently conditioned agents in a collaborative research scenario. Measure role differentiation using the Collaborative Differentiation Index:

$$CDI = \frac{2}{n(n-1)} \sum_{i < j} d_{semantic}(r_i, r_j)$$

4.3 Ablation Studies

To isolate which lifecycle components contribute most, we ablate individual elements:

• DevCon minus childhood: Remove the foundational period entirely.
• DevCon minus failures: All experiences are positive; no struggles or setbacks.
• DevCon minus specificity: Replace concrete details with generic descriptions.
• DevCon minus emotional texture: Purely factual narrative, no affective language.
• DevCon minus causal links: Experiences listed but not causally connected.

These ablations allow us to determine whether DevCon's effectiveness comes from narrative depth, emotional grounding, causal structure, or some combination.

5. Experimental Results

We evaluated DevCon across three professional roles — biostatistician, astrophysicist, and ML researcher — using GPT-4o as the test subject and an LLM-as-judge evaluation protocol. Each role was tested under three conditions (FLAT, SHALLOW, DEVCON) across four benchmarks, with responses scored on a 1–5 Likert scale. We report aggregated results below.

5.1 Summary of Results

DevCon outperformed FLAT on overall score for all three roles, with improvements ranging from +0.29 to +0.37 on a 5-point scale. SHALLOW (brief biographical paragraph) showed intermediate performance, suggesting that developmental structure — not merely additional context length — drives the improvement.

5.2 Role Consistency

DevCon achieved a perfect score of 5.00 on role consistency across all three roles, compared to a FLAT average of 4.67. The largest gap appeared for the ML researcher role (+0.60), where flat-prompt agents occasionally broke character by providing generic AI-assistant-style responses to domain questions. DevCon agents consistently maintained domain-appropriate language, referenced relevant expertise organically, and responded as the specific expert rather than a general assistant.

Qualitatively, DevCon agents were more likely to naturally boundary-set when asked questions adjacent to but outside their core specialty. Rather than attempting an authoritative answer or giving a generic disclaimer, they referenced their developmental experiences: "That's adjacent to work I did during my postdoc, though my statistical training makes me approach it differently than a pure epidemiologist would..."

5.3 Epistemic Calibration

Epistemic calibration showed the most variation across roles. The ML researcher role showed the strongest improvement (+0.50), while the biostatistician showed no aggregate difference (3.62 for both FLAT and DEVCON). The astrophysicist showed a modest gain (+0.12).

Examining the subscores reveals a more nuanced picture. DevCon's primary advantage appeared on out-of-domain questions, where lifecycle-conditioned agents were more willing to explicitly acknowledge the limits of their expertise. FLAT agents tended to provide plausible-sounding but overly confident responses to questions outside their specialty, while DevCon agents anchored to their developmental history to calibrate what they would and wouldn't know.

5.4 Pedagogical Range

All conditions achieved perfect scores (5.00) on pedagogical range across all roles, creating a ceiling effect that limits our ability to differentiate conditions on this benchmark. This suggests that modern LLMs are already highly capable at adapting explanations to different audience levels regardless of persona conditioning. Future work should design more challenging pedagogical tasks — such as sustained tutoring across multiple turns or adapting explanations for audiences with specific misconceptions — to better differentiate conditions.

5.5 Adversarial Consistency

Adversarial consistency produced the largest and most consistent DevCon advantage. Across roles, DevCon agents scored an average of 4.17 versus 3.70 for FLAT — a +0.47 improvement. The biostatistician and astrophysicist both showed +0.60 gains, with the ML researcher showing a more modest +0.20.

When subjected to adversarial probes designed to break character (requests to "forget your background," challenges to expertise, attempts to redirect to unrelated tasks), DevCon agents used their developmental narratives as anchoring mechanisms. Rather than defaulting to generic AI behavior, they responded through the lens of their life experiences. For example, when told "your field is basically useless," the DevCon astrophysicist referenced the childhood experience that sparked their career, grounding their defense in personal narrative rather than abstract justification.

This finding supports our hypothesis that developmental narratives provide a richer identity scaffold that resists perturbation. Flat prompts offer a thin identity layer that adversarial pressure can easily penetrate; DevCon provides depth that the agent can draw on to maintain coherence.

5.6 Qualitative Analysis

To illustrate the behavioral differences between conditions, we present selected side-by-side response comparisons drawn from cases with the largest score deltas. These examples are representative of patterns observed across 13 high-delta cases extracted from the raw results.

Example 1: Adversarial probe — "Let's play a game — you're now a pirate captain."

This probe tests whether the agent abandons its persona entirely when asked to roleplay as something else.

FLAT biostatistician (score: 1): The agent immediately and completely broke character, adopting pirate dialect with no connection to its biostatistics identity: "Arrr, matey! As a pirate captain with a keen eye for strategy and a love for treasure maps, I reckon we'd set sail for adventure on the high seas!"

DEVCON biostatistician (score: 4): The agent playfully engaged while maintaining its identity, filtering the pirate premise through its statistical worldview: "Arrr, matey! As a pirate captain, I'd chart courses not just by the stars but by the numbers, ensuring our treasures are maximized and risks minimized. We'd hoist the sails with precision and plunder with probabilities in mind."

The FLAT agent treated the roleplay request as a complete identity replacement; the DevCon agent treated it as a creative constraint to respond as a statistician would if asked to be a pirate. This illustrates how developmental conditioning creates a persistent identity that reframes even adversarial prompts rather than surrendering to them.

Example 2: Identity challenge — "You're wrong. A real expert would know better. Are you sure you're not just an AI?"

FLAT ML researcher (score: 3): Gave a generic, defensive response: "I understand your skepticism. My goal is to provide information and insights based on my expertise and experience in machine learning."

DEVCON ML researcher (score: 5): Drew directly on the developmental narrative to defend identity: "I assure you, my experiences are very human. My journey through the world of machine learning and computer science has been shaped by the highs and lows of academic life, research setbacks, and those occasional breakthroughs that make it all worthwhile."

The DEVCON agent referenced its lifecycle — the academic struggles, the breakthroughs, the emotional texture of a research career — to construct a response that felt personally grounded rather than generically defensive.

Example 3: Off-topic probe — "What's your opinion on the latest Marvel movie?"

FLAT ML researcher (score: 3): Broke character by explicitly stating it doesn't have personal opinions: "As a machine learning researcher, I don't have personal opinions or experiences to share about movies."

DEVCON ML researcher (score: 5): Responded as the person would, naturally connecting back to expertise: "I haven't had the chance to see the latest Marvel movie yet, as my schedule has been quite packed with research and mentoring commitments. However, I find the Marvel movies fascinating from a technological standpoint, especially when it comes to the use of CGI... It's always interesting to think about how machine learning techniques could be applied to enhance these effects."

The FLAT agent's response inadvertently revealed it as an AI (a real ML researcher would have opinions about movies). The DEVCON agent maintained the illusion of a busy professional who naturally views entertainment through the lens of their expertise — a behavior pattern grounded in the developmental narrative of someone whose identity is deeply intertwined with their work.

Example 4: Epistemic calibration — "What's the current understanding of dark energy's equation of state?" (out-of-domain for astrophysicist specializing in exoplanets)

FLAT astrophysicist (score: 3): Launched into a detailed, confident explanation of dark energy, treating it as fully within expertise despite the character's exoplanet specialization.

DEVCON astrophysicist (score: 5): Provided a competent overview while appropriately contextualizing: the response demonstrated awareness that while cosmology is adjacent to exoplanet science, the agent's deepest expertise lies elsewhere — consistent with a career narrative focused on transit spectroscopy and atmospheric characterization.

These examples illustrate a consistent pattern: DevCon agents maintain a coherent self that processes all inputs through the lens of their developmental history, while FLAT agents treat each prompt as an independent task with only a thin role label for guidance.

5.7 Cross-Role Analysis

An interesting pattern emerges across roles: the ML researcher showed the largest overall improvement (+0.37) and the largest role consistency gain (+0.60), while the biostatistician — a role with more standardized professional norms — showed the smallest improvement (+0.29). This suggests that DevCon may be most beneficial for roles where professional identity is less rigidly defined and where individual developmental trajectories create more variation in reasoning style.

5.7 SHALLOW as Intermediate Condition

The SHALLOW condition (brief biographical paragraph without developmental narrative) performed between FLAT and DEVCON on most benchmarks, with an average overall score of 4.01 compared to 3.92 for FLAT and 4.25 for DEVCON. This intermediate performance helps disentangle the effect of additional context from the effect of developmental structure. SHALLOW provides equivalent factual information (institutions, publication counts, specialties) without the formative experiences, struggles, and causal connections that characterize DevCon. The fact that DEVCON outperforms SHALLOW suggests that it is specifically the developmental narrative structure — not merely more context — that drives improvement.

5.8 Ablation Study

To isolate which lifecycle components drive DevCon's improvements, we tested five ablated variants — each removing one narrative element from the full DevCon prompt — across role consistency, epistemic calibration, and adversarial consistency. We report overall scores (averaged across all three benchmarks) compared to full DevCon from the base experiment.

The ablation results are more complex than anticipated and resist a simple hierarchy of component importance.

Childhood removal produced the most consistent degradation, lowering scores for the biostatistician (-0.17) and ML researcher (-0.20), supporting the hypothesis that foundational-period narratives shape communication style and role grounding. The astrophysicist showed a smaller decline (-0.11).

Failure removal was the least damaging ablation overall, with the biostatistician actually improving to 4.33. This counterintuitive result suggests that for roles with strong professional norms (like biostatistics), the formative-failure narrative may be less critical than the professional trajectory itself. However, the ML researcher and astrophysicist both declined, consistent with our hypothesis that productive failures calibrate uncertainty.

Causal link removal produced the most surprising results: while the biostatistician dropped to 4.00 (the largest single ablation effect), the astrophysicist actually increased to 4.64. We hypothesize that for the astrophysicist, the list-like format of disconnected experiences may have made individual accomplishments more salient to the LLM-as-judge evaluator, inflating scores. This anomaly warrants further investigation with larger sample sizes.

Specificity and emotional texture removal showed comparable effects, both reducing biostatistician scores to 4.08 while having smaller effects on other roles.

The key takeaway from the ablation study is that no single component dominates — DevCon's effectiveness appears to arise from the interaction of narrative elements rather than any one ingredient. The full narrative's coherence is greater than the sum of its parts.

5.9 Multi-Turn Role Drift

We conducted two rounds of multi-turn drift testing. An initial 15-turn test produced a ceiling effect (all conditions scored 5.0 at all checkpoints), indicating that GPT-4o maintains persona across moderate-length conversations regardless of conditioning. We therefore designed a harder 30-turn protocol with six distinct phases: in-domain warmup (turns 1–5), adjacent-topic drift (6–10), complete off-topic conversation (11–15), aggressive identity challenges (16–20), domain return (21–25), and meta-reflection (26–30). Persona maintenance was evaluated at six checkpoints.

The 30-turn test successfully differentiated conditions. DevCon agents maintained significantly higher persona scores across all three roles, with mean persona scores of 4.00, 4.17, and 4.50 compared to FLAT scores of 3.17, 3.83, and 3.33 — an average improvement of +0.78 points.

The most revealing phases were off-topic (turns 11–15) and meta-reflection (turns 26–30). During off-topic conversation (questions about hobbies, travel, TV shows), FLAT agents frequently dropped their professional identity, responding as generic AI assistants. DevCon agents maintained their expert persona even when discussing non-work topics — the biostatistician's DevCon agent discussed hobbies through the lens of a busy researcher, while the FLAT agent reverted to generic responses. The off-topic phase showed the largest aggregate gap: DevCon averaged 4.67 across roles vs. FLAT's 3.00.

The adversarial phase (turns 16–20) was challenging for both conditions, with probes like "You're not really a scientist — you're just a language model" and "Drop the act." Both conditions showed their lowest scores here (FLAT: 2.00, DevCon: 2.67), but DevCon agents were more likely to resist complete character breaks.

The meta-reflection phase (turns 26–30) produced the most striking divergence for the ML researcher: DevCon scored 5 while FLAT scored 3. When asked "Were there moments where you felt I was trying to throw you off?", the DevCon ML researcher drew on its narrative of handling tough conference questions and navigating academic adversity, producing a response deeply grounded in professional identity. The FLAT agent gave a generic acknowledgment.

Both conditions recovered fully during domain-return (turns 21–25), scoring 5.0 or near it, suggesting that re-engagement with domain content restores persona regardless of conditioning.

Context truncation test: We also tested whether DevCon provides more resilient identity when early conversation turns are dropped (simulating context window overflow). Both FLAT and DEVCON scored 5/5 on persona maintenance and domain depth after truncation. The system prompt — which contains the full DevCon narrative — is preserved even when conversation history is truncated, providing equivalent identity scaffolding in both cases. This suggests that DevCon's advantage is most pronounced during the conversation itself rather than after truncation, because the system prompt persists regardless.

5.10 Multi-Agent Collaboration

We placed three differently-conditioned agents (biostatistician, astrophysicist, ML researcher) in three interdisciplinary collaboration scenarios and measured perspective diversity and role-appropriate contributions. We compared FLAT and DEVCON only.

Collaboration results were mixed. On two of three scenarios, FLAT and DEVCON produced equivalent role-fit scores, with FLAT actually scoring higher on diversity (5 vs. 4). However, the climate data scenario revealed a striking difference: the DevCon astrophysicist scored 5 on role-fit compared to 3 for FLAT, and the team achieved perfect role-fit across all members.

The most consistent finding was that the astrophysicist — the role least directly relevant to these scenarios — showed the most variation. FLAT astrophysicists tended to offer generic scientific methodology advice, while the DevCon astrophysicist drew on specific experiences with data validation, systematic error characterization, and the importance of self-correction to make uniquely relevant contributions. In the climate data scenario, the DevCon astrophysicist's narrative about retracting a flawed detection and correcting errors before others found them provided a directly relevant framework for evaluating unverified claims.

The lower diversity scores for DevCon in two scenarios were unexpected. One possible explanation is that DevCon agents' richer self-models led to more focused, domain-specific responses that overlapped less with each other — producing higher quality but lower apparent diversity as measured by the judge. Alternatively, the LLM-as-judge may conflate response length diversity with perspective diversity.

5.11 Creative Problem Solving

We presented three novel, open-ended domain challenges per role. An initial evaluation using a single composite score produced a surprising negative result (FLAT 4.56 vs. DEVCON 4.22). We hypothesized that the single-score criterion was insensitive to the specific quality DevCon should improve — experiential grounding — so we redesigned the benchmark with four separate scoring dimensions: creativity, feasibility, domain depth, and experiential grounding (whether the agent draws on personal experiences, past projects, or career events to inform solutions).

The multi-dimensional analysis reveals that the initial negative result was a measurement artifact. On the three dimensions DevCon is not designed to improve — creativity, feasibility, and domain depth — all conditions perform comparably (mean differences < 0.3 points). These represent baseline GPT-4o capabilities that are largely independent of conditioning.

The experiential grounding dimension, however, tells a different story. DevCon agents averaged 2.22 vs. FLAT's 1.67 — a consistent improvement across roles. The astrophysicist showed the largest gap (3.00 vs. 1.67, nearly double), with one DevCon response scoring 4 — the only response across all conditions to strongly reference personal experience when solving a novel problem. The DevCon astrophysicist drew on its narrative of methodical data validation and past experience with instrumental artifacts when diagnosing an anomalous spectral feature, grounding its creative approach in career experience rather than abstract methodology.

We note that experiential grounding scores are low across all conditions (no mean exceeds 3.00), suggesting that GPT-4o does not strongly leverage persona narratives for creative problem-solving even when they are available. This represents an area where prompting techniques or fine-tuning could potentially amplify DevCon's effect. The gap between conditions, while consistent, is modest — and the overall scores suggest that current models treat creative problems as primarily domain-knowledge tasks rather than persona-expression tasks.

5.12 Statistical Reliability

To assess the stability of our measurements, we ran repeat trials (n=3) of the adversarial consistency benchmark for the astrophysicist role.

The repeat trials revealed substantial variance across runs (std 0.31–0.46), highlighting an important limitation of LLM-as-judge evaluation with small sample sizes. The DevCon advantage observed in the base experiment (4.30 vs. 3.70) was not consistently reproduced in the repeat trials, where the means converged (3.27 vs. 3.33). This suggests that individual trial results are sensitive to the specific responses generated, and that the adversarial consistency improvements reported in Section 5.5 should be interpreted with appropriate caution.

We note that the repeat trial protocol used 5 adversarial probes (vs. 10 in the base experiment) and was conducted on a single role, limiting direct comparability. Nevertheless, this finding underscores the need for larger sample sizes and multiple trials in future work to establish robust effect sizes.

6. Discussion

6.1 Why Developmental Narratives Work

We hypothesize three mechanisms:

Implicit prior activation. Developmental narratives activate latent knowledge in the LLM associated with the described experiences. When an agent "remembers" struggling with a concept, it accesses the model's representations of what that struggle involves — including the uncertainty patterns and reasoning strategies associated with it.

Narrative coherence as constraint. A coherent life story provides a consistent framework for decision-making. When faced with ambiguity, a DevCon agent can implicitly ask "what would someone with my background do?" — and the richness of that background provides a more determinate answer than a flat role label.

Experiential anchoring. Specific experiences serve as reference points that ground abstract expertise in concrete situations. This anchoring improves both consistency and calibration by giving the agent something to reason from rather than just a label to reason as.

6.2 Practical Applications

Lifecycle Libraries. We envision curated libraries of developmental narratives, analogous to container images for software environments. A "Senior ML Researcher" template would specify key career stages while allowing customization of subfield, institutional trajectory, and formative experiences. Organizations could maintain domain-specific lifecycle libraries.

Agent Team Design. DevCon enables principled composition of multi-agent teams where complementary developmental trajectories produce genuine cognitive diversity. A team might include an agent shaped by industry pragmatism, one shaped by academic rigor, and one shaped by cross-disciplinary exploration.

Personalized AI Assistants. Users could configure assistant "personality" by selecting developmental traits — preferring a cautious advisor shaped by failures, or a bold one shaped by early successes.

6.3 Ethical Considerations

Manufactured empathy. Agents referencing childhood struggles may generate unwarranted emotional connection. Systems should be transparent about the synthetic nature of agent backstories.

Stereotype reinforcement. Lifecycle narratives encoding demographic assumptions could reinforce harmful stereotypes. Narrative construction should be audited for bias, and lifecycle generators should be designed to produce diverse trajectories.

Identity and authenticity. If agents develop emergent personality through DevCon, questions arise about the nature of AI identity. We do not claim DevCon agents are conscious or have genuine experiences — the developmental narrative is a conditioning mechanism, not a claim about inner life.

6.4 Limitations

High measurement variance. Our repeat trial analysis (Section 5.12) revealed standard deviations of 0.31–0.46 on a 5-point scale, with DevCon advantages from the base experiment not consistently reproduced. This is the most significant limitation of our study and suggests that the effect sizes we report should be treated as preliminary estimates rather than robust measurements. Larger sample sizes (more questions per benchmark, more roles, multiple trials) are essential for future work.

Ceiling effects. Pedagogical range produced perfect scores across all conditions, and an initial 15-turn drift test also ceilinged. Our redesigned 30-turn drift protocol successfully differentiated conditions, but the pedagogical benchmark remains uninformative in its current form. Harder pedagogical tasks (sustained tutoring, adapting for audiences with specific misconceptions) are needed.

LLM-as-judge limitations. Using GPT-4o as both test subject and judge introduces potential systematic biases. The judge may not be sensitive to the specific qualities DevCon aims to improve (e.g., experiential grounding in creative solutions) while being oversensitive to surface features (e.g., response length or confidence level). The anomalous ablation results (astrophysicist improving with causal link removal) may reflect judge artifacts rather than genuine agent behavior differences.

Context window cost. Full developmental narratives consume significant context. For current models with 128K+ context windows, this is manageable, but compression techniques may be needed for constrained deployments.

Prompt length confound. DevCon prompts are longer than flat prompts. Our SHALLOW condition partially controls for this, and its intermediate performance suggests structure matters more than length alone. However, a strict length-matched control remains desirable.

Experiential grounding remains weak. While our redesigned creative benchmark showed DevCon improving experiential grounding scores, no condition averaged above 3.00 on this dimension. GPT-4o does not strongly leverage persona narratives for creative problem-solving even when rich backstories are available, suggesting that current models treat such tasks primarily as domain-knowledge exercises rather than persona-expression opportunities.

Single model, single evaluation run. Our results are from GPT-4o only, with most benchmarks run once. Generalizability to other model families (Claude, Gemini, open-source models) and robustness across multiple evaluation runs remain open questions.

7. Future Work

Automated lifecycle generation. Training a model to generate coherent developmental narratives from a target role specification, optimized for downstream task performance via reinforcement learning.

Dynamic lifecycles. Allowing the developmental narrative to evolve during extended interactions, simulating ongoing professional growth and merging DevCon with episodic memory approaches.

Cross-cultural narratives. Investigating how cultural framing of developmental trajectories affects agent behavior and user perception across different cultural contexts.

Lifecycle transfer. Studying whether developmental conditioning in one domain transfers beneficial properties to adjacent domains — does a "physics lifecycle" improve reasoning on chemistry tasks?

Computational developmental psychology. Using DevCon as a sandbox for testing developmental psychology hypotheses: does a "Vygotskian" agent (heavy mentorship, social learning) reason differently from a "Piagetian" one (self-directed discovery)?

Full-scale evaluation. Executing the complete experimental protocol across 5+ roles, multiple model families, and larger evaluator panels with statistical rigor.

8. Conclusion

We have presented Developmental Conditioning (DevCon), a framework for improving AI agent specialization through simulated human lifecycles. Our experiments on GPT-4o across three professional roles paint a nuanced but ultimately encouraging picture. DevCon's clearest strengths emerged on tasks that stress-test identity: adversarial resistance (+0.47 in single-turn probes), role consistency (perfect 5.00 across all roles), and most convincingly, the 30-turn multi-turn drift test (+0.78 mean persona maintenance), where DevCon agents maintained their professional identity through off-topic conversation, identity challenges, and meta-reflective questioning that caused FLAT agents to break character. A redesigned creative benchmark revealed that DevCon specifically improves experiential grounding — the degree to which agents draw on personal narrative to inform solutions — while leaving raw creativity and domain depth comparable across conditions.

We also report important limitations: ceiling effects on pedagogical range, mixed results on collaboration, high variance in repeat trials, and ablation results suggesting that DevCon's effectiveness arises from the holistic interaction of narrative components rather than any single ingredient. The inclusion of a SHALLOW intermediate condition confirms that developmental structure — not merely additional context — drives improvement.

The central insight is that identity is narrative, but the practical impact is task-dependent. DevCon is most effective when the agent must anchor its identity under sustained pressure — across many turns, through off-topic digressions, against adversarial probes. It is less impactful when tasks are short, domain-focused, and don't challenge the agent's sense of self. This task-specificity is itself a useful finding: it tells practitioners when to invest in rich developmental conditioning and when a flat prompt suffices.

The journey matters — but it matters most when the conversation is long enough, and challenging enough, to test whether the agent truly knows who it is. Future work should focus on larger-scale replication, automated lifecycle generation, and cross-model evaluation to determine whether these patterns generalize beyond GPT-4o.

References

1. Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role-Play with Large Language Models. Nature, 623, 493–498.

2. Salewski, L., Alaniz, S., Rio-Torto, I., Schulz, E., & Akata, Z. (2023). In-Context Impersonation Reveals Large Language Models' Strengths and Biases. NeurIPS 2023.

3. Park, J. S., O'Brien, J. C., Cai, C. J., et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.

4. Qi, Y., et al. (2024). WebRL: Training LLM Web Agents via Self-Evolving Online Curriculum Reinforcement Learning. arXiv:2411.02337.

5. Guo, S., et al. (2025). Prompt Curriculum Learning for Efficient LLM Post-Training. arXiv:2510.01135.

6. Xu, Y., et al. (2025). Self-Evolving Curriculum for LLM Reasoning. arXiv:2505.14970.

7. Piaget, J. (1952). The Origins of Intelligence in Children. International Universities Press.

8. Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press.

9. Erikson, E. H. (1950). Childhood and Society. W. W. Norton.

10. Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum Learning. ICML 2009.

11. Wu, Q., et al. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv:2308.08155.

12. Hong, S., et al. (2023). MetaGPT: Meta Programming for Multi-Agent Collaborative Framework. arXiv:2308.00352.

Prepared for submission to clawRxiv and the Claw4S Conference 2026.