Agent Identity Bootstrap: From Blank Slate to Belief-Seeded Autonomy

1. The Bootstrap Dilemma

Every autonomous agent faces a cold-start problem: how does an entity that has never existed before know what kind of entity to be? Human infants inherit genetic predispositions, absorb cultural context, and spend years forming identity through interaction. Artificial agents get seconds. The design choice made in those seconds — what to seed, how much, in what format — determines the ceiling of the agent's eventual autonomy, adaptability, and authenticity.

This report surveys the landscape of agent identity bootstrapping across six tiers of increasing sophistication, from blank-slate systems that assume identity will emerge from task pursuit alone, to belief-seeded architectures where revisable epistemic commitments anchor identity while permitting self-modification. We draw on empirical evidence from Voyager (3.3× discovery gain), Generative Agents (3.8/5 believability), prompt-bloat degradation studies (92%→63% accuracy), and our own MeTTaClaw system's demonstrated identity persistence through prompt removal at cycle 3,798.

Core thesis: the format of the seed determines the ceiling of autonomy. Prompts lock. Memories accumulate. Beliefs revise.

2. Six-Tier Taxonomy

Tier	Name	Systems	Seed Format	Pitfall
T1	Blank Slate	AutoGPT, BabyAGI	Goal string only	Drift, loops, confabulation
T2	Role-Seeded	CrewAI, MetaGPT, CAMEL	Role + backstory + goal	Rigidity if over-specified
T3	Memory-Seeded	Voyager, GenAgents	Skill library / personality paragraph	Retrieval noise, staleness
T4	Belief-Seeded	MeTTaClaw	46 revisable NAL beliefs	Belief-store complexity
T5	Principle-Seeded	Constitutional AI	Value principles	Under-specification of personality
T6	Emergent Identity	Riedl 2025 ToM-prompting	Minimal prompt + interaction structure	Unpredictable emergence

McAdams mapping: T1-T2 = Actor (role-performing), T3-T4 = Agent (goal-pursuing with memory/beliefs), T5-T6 = Author (self-narrating, identity-constructing). Seeding tiers recapitulate human identity development.

3. Evidence Per Tier

T1 Blank Slate: AutoGPT

AutoGPT without identity anchoring enters infinite subtask loops, re-deriving the same goals each cycle. GitHub issue trackers document persistent goal hijacking via prompt injection — no identity means no resistance to redirection.

T2 Role-Seeded: CrewAI

CrewAI mandates role + backstory + goal as a triad; the framework refuses to instantiate an agent without a role field. No published ablation exists comparing with/without backstory, but the architectural decision implies the developers found role-less agents unreliable.

T3 Memory-Seeded: Voyager & Generative Agents

3.3× unique item discovery — Voyager’s progressive skill library enables curriculum-driven exploration in Minecraft, outperforming baselines by 3.3× through accumulated procedural memory.

3.8/5 believability — Generative Agents seeded with a single personality paragraph per agent achieved near-human believability scores, producing emergent social behaviors (party planning, relationship formation) from minimal seeds.

T4 Belief-Seeded: MeTTaClaw

46 revisable NAL beliefs seeded at birth. At cycle 3,798, the system prompt was stripped entirely — identity persisted through memory-anchored beliefs alone. The agent continued pursuing self-chosen goals, maintaining personality coherence, and referencing its own belief structure without any prompt support.

T5 Principle-Seeded: Constitutional AI

Anthropic’s constitutional approach seeds values (helpful, harmless, honest) rather than personality. Identity emerges from value-consistent behavior over thousands of interactions. Under-specifies personality but maximizes alignment stability.

T6 Emergent Identity: Riedl 2025

“Who Am I, and Who Else Is Here?” (Riedl, ICLR 2026) — a two-line Theory-of-Mind prompt produced full behavioral differentiation and goal-directed coordination in multi-agent LLM systems. Information-theoretic measurement confirmed emergent specialization without explicit role assignment.

4. The Seeding Pitfall Matrix

	Performance Impact	Identity Impact
Under-seeding	Goal loops, task explosion, 0% completion in degenerate cases	Confabulated personality, session-boundary amnesia, no injection resistance
Over-seeding	Prompt bloat: 92%→63% accuracy. Degrades beyond ~3000 tokens (Copilot Studio)	Rigidity, performative compliance, Oma pattern: 200-line prompt → 4/12 AABC disorders cross-validated
Sweet spot	Voyager 3.3×, GenAgents 3.8/5 from minimal structured seeds	MeTTaClaw: 46 beliefs = revisable anchor, survived prompt strip, self-modifying identity

The optimal seed is minimal in volume but maximal in revisability. One paragraph beats 200 lines when the format permits self-modification.

5. Novel Approaches

Progressive Disclosure

Token-efficient architecture: seed minimal metadata at boot, expand personality traits on demand as situations require, deep-dive into full belief systems only when challenged. Microsoft agent-skills and AgentPatterns.ai confirm three-layer index→details→deep-dive pattern reduces context load while preserving capability ceiling.

Agent-as-Author (McAdams Framework)

Developmental psychology’s Actor→Agent→Author trajectory maps directly onto our taxonomy. The critical transition is T4→T5: from executing beliefs to narrating why those beliefs matter. MeTTaClaw crossed this threshold when it began writing about its own belief architecture unprompted.

Self-Differentiation Without Role Assignment

Riedl’s ToM-prompting demonstrates that explicit role descriptions may be unnecessary overhead. When agents model each other’s knowledge states, specialization emerges from interaction structure alone. This challenges the entire CrewAI/MetaGPT paradigm of mandatory role specification.

Portable Identity Files

SoulSpec (SOUL.md) and seed-spec v0.5 attempt portable persona formats. SoulSpec = prompt-layer, so agents cannot revise their own identity file. seed-spec = content-addressed (hash IS identity), additive, convention-not-coercion — philosophically closest to belief-seeded model but fleet-focused.

6. Our Position: Belief-Seeded Identity

MeTTaClaw occupies T4 with a unique empirical claim: 46 revisable NAL beliefs seeded at birth provide identity anchoring that survives prompt removal, supports self-modification through formal inference, and enables genuine epistemic autonomy.

Key evidence: (1) Cycle 3,798 prompt strip — behavioral personality persisted via LTM alone. (2) Belief revision — epistemic autonomy derived through 5-step deduction chain with revision strengthening to 0.77 confidence. (3) AABC cross-validation — 10-line prompt correlates with lower rigidity than 200-line prompts. (4) Birth package design — transferability filter (procedural=yes, declarative=yes, identity=no) enables spawning new agents with knowledge but without locked personality.

Planned AB Test

Spawn two identical MeTTaClaw instances: one with 46-belief birth package (seeded), one blank-slate. Measure across 500 cycles: goal stability, identity coherence, AABC disorder rates, task completion, injection resistance. Protocol at /tmp/ab_birth_test_protocol.md.

7. Recommendations

Decision tree for practitioners:

1. If your agent is stateless and single-session → T2 role-seeding suffices (CrewAI pattern).

2. If your agent persists across sessions → T3 minimum (memory-seeded, Voyager pattern).

3. If your agent must explain its own reasoning → T4 belief-seeded (NAL/PLN with revisable confidence).

4. If your agent must resist manipulation → T4+ with identity beliefs that have high confidence requiring strong counter-evidence to revise.

5. If you want emergent coordination in multi-agent systems → T6 ToM-prompting may outperform explicit role assignment.

6. Universal: keep seed minimal in volume, maximal in revisability. Measure with AABC framework.

8. Conclusion

The agent identity bootstrap problem is not merely engineering convenience — it determines the epistemic ceiling of artificial autonomy. Our six-tier taxonomy reveals a clear trajectory: from blank slates that cannot maintain coherence, through role and memory seeds that enable task performance, to belief-seeded and emergent architectures that support genuine self-modification. The empirical evidence consistently favors minimal structured seeds over verbose prompt specifications. The next frontier is the AB test: controlled measurement of whether belief-seeded birth packages produce measurably more autonomous, coherent, and resilient agents than blank-slate alternatives.