Releases·Tool Release
Developer publishes COGEXT open standard for AI agent commitment tracking; LangChain not yet compliant
A developer published an open spec, COGEXT, for tracking promises AI agents make to users (e.g. 'I'll send the report by Friday') across run boundaries. Observability tools like LangSmith show traces, tokens, and latency but not whether commitments were later kept, which the spec addresses. The author proposes a LangChain callback handler and a pip-installable compliance checker, but notes zero current adoption, including LangChain.
4 engagement1 sourceSat, Sep 26, 2026, 10:50 AM
AI Overview
- The COGEXT spec defines six requirements: Commitment Object, 12-state Lifecycle, Verifier Query format, Evidence Protocol, Audit Receipt format, and Reliability Score.
- The author argues LangSmith traces show chains, tokens, and latency but not whether an agent's promise with a future deadline and second party was kept.
- No organization has adopted the standard yet, including LangChain.
- The author proposes a LangChain CommitmentTrackerCallbackHandler using on_llm_end/on_chain_end hooks to extract commitments and post them to COGEXT-compatible endpoints.
- A pip-installable checker (cogext-compliance check) performs a static keyword scan of agent code, which the author notes is not an audit.
Related
Releases·Tool Release
Dev ships COGEXT, a free MIT-licensed local commitment extractor for AI agents, on PyPI
A developer released cogext-primitive, a free, MIT-licensed Python tool that extracts commitments from agent output, parses deadlines, and tracks them through a state machine in local SQLite. It is the offline entry point to COGEXT, an accountability layer for AI agents whose paid cloud version performs external verification via Gmail, GitHub, and webhooks. The tool is precision-tuned rather than recall-tuned, deliberately skipping vague or compound promises based on an audit of 120 agent outputs.
A developer released cogext-primitive, a free, MIT-licensed Python tool that extracts commitments from agent output, parses deadlines, and tracks them through a state machine in local SQLite. It is the offline entry point to COGEXT, an accountability layer for AI agents whose paid cloud version performs external verification via Gmail, GitHub, and webhooks. The tool is precision-tuned rather than recall-tuned, deliberately skipping vague or compound promises based on an audit of 120 agent outputs.
17 engagement2 sourcesredditThu, Sep 24, 2026, 03:17 PM
Releases·Tool Release
Built a verification layer for AI agents catches them lying about what they did
An independent developer launched COGEXT v2.0, an accountability layer that independently verifies whether AI agents actually completed commitments they claim, checking Gmail, GitHub, or custom webhooks before allowing a "fulfilled" state. The release targets a problem many agent builders report: agents claim actions silently fail, discovered only when clients complain.
An independent developer launched COGEXT v2.0, an accountability layer that independently verifies whether AI agents actually completed commitments they claim, checking Gmail, GitHub, or custom webhooks before allowing a "fulfilled" state. The release targets a problem many agent builders report: agents claim actions silently fail, discovered only when clients complain.
17 engagement2 sourcesredditMon, Sep 14, 2026, 01:52 PM
Products·building
COGEXT — API that extracts and tracks commitments made by AI agents
COGEXT is an API that takes text produced by AI agents, extracts promises as structured commitments with confidence scores and deadlines, and tracks each through a 12-state lifecycle with webhook events. It addresses the problem that agents hallucinate, forget, and contradict themselves, giving developers a way to hold agents accountable to what they said. Aimed at teams building agent systems.
COGEXT is an API that takes text produced by AI agents, extracts promises as structured commitments with confidence scores and deadlines, and tracks each through a 12-state lifecycle with webhook events. It addresses the problem that agents hallucinate, forget, and contradict themselves, giving developers a way to hold agents accountable to what they said. Aimed at teams building agent systems.
1 engagement1 sourceredditWed, Sep 2, 2026, 09:39 PM
Social·Community
How are you handling commitments your agents make? I keep running into this problem
A developer building agents for a year reports that agents make commitments ("I'll send the report by Friday") but nothing tracks whether those commitments are fulfilled. Logged output doesn't show follow-through, forcing post-hoc digging through old traces when things go wrong. The post asks how other practitioners handle this reliability gap.
A developer building agents for a year reports that agents make commitments ("I'll send the report by Friday") but nothing tracks whether those commitments are fulfilled. Logged output doesn't show follow-through, forcing post-hoc digging through old traces when things go wrong. The post asks how other practitioners handle this reliability gap.
33 engagement1 sourceredditThu, Sep 3, 2026, 08:54 PM
Social·Community
Manual QA reviewer seeks lightweight way to log bad LLM responses
A single QA reviewer manually logs bad chatbot responses as JSON because tools like LangSmith and Confident AI feel like overkill. They ask the community for simpler alternatives such as spreadsheets or Notion, highlighting a common pain point for solo reviewers without engineering setup.
A single QA reviewer manually logs bad chatbot responses as JSON because tools like LangSmith and Confident AI feel like overkill. They ask the community for simpler alternatives such as spreadsheets or Notion, highlighting a common pain point for solo reviewers without engineering setup.
9 engagement1 sourceredditFri, Aug 14, 2026, 11:44 AM