code-agents

A living collection of frontier research on code agents, from the code we build to the worlds we act in.

Turn feature specs into merged PRs with a self-supervising swarm of coding agents — parallel execution, isolated sandboxes, DAG dependencies. Open-source, self-hostable, model-agnostic (Claude / Gemini / Codex).

[ICML 2026] JustAsk: Curious Code Agents Reveal System Prompts in Frontier LLMs | Verified on Claude Code | Autoresearch for System Prompt Extraction

Agent skill for generating professional PowerPoint decks. 7 composable skills and a purpose-built CLI.

A service-boundary-aware document exchange center for coordinating heterogeneous LLM code agents via MCP. Implements versioned Markdown store, pub-sub notifications, and diff-aware update protocol.

CodeAgents: A Token-Efficient Framework for Codified Multi-Agent Reasoning in LLMs

Deterministic method-first retrieval for AI coding agents.

Apply unified diffs and fuzzy search/replace edits in pure Python - the patch engine for code agents. No git, no patch binary.

Security-first control plane for autonomous AI code agents: sandboxed execution, hash grounding, diff validation, verification, and full auditability. The governance layer agents need before they touch production.

Scaffolds reproducible coding-agent RL task environments from real PR fixes: content-anchored diff inversion, F2P/P2P harness generation verified by pytest subprocesses, solution-leak detection, and byte-identical bundles proven across a 3-OS CI matrix.

Autonomous code review & remediation engine with LibCST patching, event-sourced FSM orchestration, sandboxed verification, and authenticated human approval.

Issue localization baselines (BM25 + code graph expansion) for the Gemma 4 Developer Agent competition

Loop Engineering for automated repair of visual speech recognition inference pipelines.

Code-agent safety on RedCode-Gen: retrieval + reflection raise Qwen2.5-7B refusals from 1.9% to 38.1%

Local-first graph-aware context engine for code agents.