ktech7moon/agent-habitat

agent-habitat is an audit-grade multi-agent orchestration framework designed to enforce fabrication-resistance as a validated contract rather than a hope. It’s aimed at teams and environments (e.g., fintech, compliance, research ops) that require provable, re-runnable evidence of what agents did and won’t accept “we don’t know” answers. Key features: - A five-agent lead-enrichment crew with persistent state, human approval checkpoint, and per-LLM-call cost telemetry. - A critic that decomposes every drafter claim and runs a five-hop pure-Python substring grounding chain against upstream evidence; failures trigger a bounded retry and then halt-not-ship. - Audit-grade cold storage (SQLite events, JSONL telemetry, LangGraph checkpoints) so every decision and LLM call is queryable without an LLM. - Two-layer retry policy (LLM infra retries + orchestrator fabrication retries), operator-editable ICP rubric (TOML), and three-tier model routing with conservative defaults. - Live calibration logs demonstrating real-world halts and cost profiles.

このRepoのdescription / GitHub topics / READMEのAI要約に出現したAI関連キーワードをカテゴリ別に表示。各バッジは該当カテゴリの詳細ランキングへリンクします。