[Paper] Bootstrapping Coding Agents: The Specification Is the Program

Source: arXiv - 2603.17399v1

Overview

Martin Monperrus demonstrates that a coding AI can bootstrap itself: given only a concise textual specification, a fresh AI agent can regenerate a correct implementation from scratch, even after the original code was produced by a different AI (Claude Code). This mirrors the classic bootstrap process used in compiler construction and highlights the specification as the true source of truth for software development.

Key Contributions

• Self‑bootstrapping demonstration: Shows a newly generated coding agent can re‑implement a 926‑word specification correctly without any hand‑written code.
• Meta‑circular insight for AI agents: Extends the Lisp concept of meta‑circularity to modern AI‑driven code generation.
• Specification‑centric paradigm: Argues that the specification, not the generated implementation, should be treated as the immutable artifact of record.
• Empirical validation: Uses Claude Code to produce an initial implementation, then a second‑generation agent to reproduce the same functionality, confirming reproducibility.

Methodology

1. Specification authoring – The author writes a detailed, human‑readable specification (≈ 926 words) describing the desired program behavior.
2. First‑pass generation – An existing coding agent (Claude Code) consumes the specification and emits a working implementation.
3. Bootstrapping agent creation – A new coding agent is prompted with the same specification and the source code of the first implementation, but is instructed to re‑implement from scratch rather than copy.
4. Evaluation – The regenerated program is compiled/run against the same test suite used for the first implementation. Functional equivalence and correctness are measured.

All steps rely on standard prompting techniques and off‑the‑shelf LLM APIs; no custom model training is required.

Results & Findings

• The bootstrapped agent produced a functionally identical program that passed all original test cases.
• Minor stylistic differences (naming, code layout) were observed, but semantic behavior remained unchanged.
• The experiment confirms that the specification alone is sufficient for an AI to reconstruct a correct implementation, even after the original code has been “forgotten.”
• This mirrors the classic compiler bootstrap: a compiler can compile its own source when provided with a minimal, trusted description.

Practical Implications

• Version‑controlled specifications: Teams can store the specification in a repository as the single source of truth; regenerated code can be produced on demand (e.g., after a breaking change in the underlying LLM).
• Rapid re‑implementation: When a coding agent is upgraded or replaced, existing projects can be re‑generated automatically from their specs, reducing migration friction.
• Security & compliance: Auditors can verify that the specification complies with policies; regenerated code can be inspected for unintended deviations.
• Continuous integration pipelines: A step can be added that re‑generates code from specs on each CI run, ensuring the implementation stays in sync with its description.
• Reduced technical debt: Since the implementation is not the immutable artifact, developers can focus on improving specifications rather than maintaining legacy code.

Limitations & Future Work

• Specification quality dependence – The approach hinges on a clear, complete, and unambiguous spec; vague or incomplete specs will lead to incorrect regenerations.
• Model variability – Different LLMs or even the same model with temperature changes can produce divergent code; reproducibility may require deterministic prompting or model fine‑tuning.
• Scalability – The study uses a relatively small (≈ 1 k‑word) spec; it remains open how well the method scales to large, multi‑module systems.
• Future directions – Investigate automated spec extraction from existing codebases, explore bootstrapping across heterogeneous languages, and develop tooling to integrate specification‑centric workflows into standard development environments.

Authors

• Martin Monperrus

Paper Information

• arXiv ID: 2603.17399v1
• Categories: cs.SE, cs.LG
• Published: March 18, 2026
• PDF: Download PDF