Research map / Multi-agent systems and games / LLM agents

BootstrapAgent: Turning Repository Setup into Reusable Agent Knowledge

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BootstrapAgent, a multi-agent pipeline that distills repository setup exploration into a verified, reusable .bootstrap contract, using Docker-based clean replay and trace-driven delta repair.

Problem

Repository bootstrapping (getting an unfamiliar repo into a usable dev state) is costly for code agents, and the resulting setup knowledge is lost after each conversation.

Approach

Persist bootstrap exploration as a repository-local, verifiable .bootstrap contract (setup, doctor, verify, repair knowledge) produced by discovery, planning, generation and trace-driven delta repair, with warm repair for speed and clean-replay cold verification plus a sanity check against verification weakening.

Key results

• BootstrapAgent achieves 197/212 (92.9%) clean-replay success vs. 179/212 (84.4%) for HerAgent, +8.5 points.
• Downstream Claude Code agents with .bootstrap reduce token usage by 25.9% and build time by 22.3% (per abstract).
• Removing trace-driven repair drops success by 69.7, 66.0, and 55.0 points on Repo2Run, ExecutionAgent, and Installamatic respectively.

Surprising findings

• Clean replay has only a small effect on aggregate success (e.g., 0.8 points on Repo2Run), since warm-state repair usually preserves reproducibility, yet it is still needed to catch warm-only artifacts.
• Warm-start ExecutionAgent total tokens dropped from 597.9K to 115.7K (0.19x), with the largest gains on that benchmark.

Limitations

• Establishes local bootstrap readiness rather than full CI reproduction; secrets, services, and long jobs are excluded.
• Generated .bootstrap files can become outdated and need maintenance as the project evolves.
• Struggles on repositories with missing or inconsistent version information, where repair can yield brittle fixes.

Experimental setup

Datasets: Repo2Run-Bench ExecutionAgent-Bench Installamatic-Bench

Benchmarks: Repo2Run-Bench ExecutionAgent-Bench Installamatic-Bench

Metrics: bootstrap success rate token usage wall-clock build time

Baselines: HerAgent ExecutionAgent Repo2Run Installamatic

Builds on: HerAgent Claude Code

Closest papers by meaning

• Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases (NeurIPS 2026 · similarity 0.90)
• NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents (ICML 2026 · similarity 0.90)
• RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations (NeurIPS 2026 · similarity 0.88)
• RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving (NeurIPS 2025 · similarity 0.88)
• GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging (AAAI 2026 · similarity 0.87)
• RepoAudit: An Autonomous LLM-Agent for Repository-Level Code Auditing (ICML 2025 · similarity 0.86)