COGEXT
A locally operated, offline AI agent commitment extraction and tracking tool that requires no API keys or cloud services. It uses a state-machine mechanism to perform structured extraction and lifecycle management of natural-language commitments made by AI agents, and is positioned as an accountability layer for AI agent workflows.
Timeline (last 90 days)
COGEXT is a project that establishes an accountability layer for AI agents, used to track commitments made by agents.
COGEXT tracks the lifecycle of each commitment through a state machine: open→due→overdue→fulfilled.
For an accountability system, the cost of incorrectly tracking a non-existent commitment is higher than missing an ambiguous one.
The author published an audit report on 120 agent output samples as the empirical basis for the precision-first strategy.
If no corresponding evidence is found in external systems, COGEXT blocks the Agent from marking the task as completed.
COGEXT v2.0 contains five core capabilities: a verification engine, a kill switch, a contradiction radar, a failure predictor, and public audit receipts.
COGEXT's Kill Switch implements human intervention through Slack's approve/reject interaction.
COGEXT's Contradiction Radar is used to detect self-contradictions between an agent's earlier and later statements.
COGEXT is a verification layer tool for authenticating AI Agent behavior, working by extracting commitments and independently verifying them in external systems.
The author states that COGEXT v2.0 passed 8/8 end-to-end (E2E) tests in a production environment.
7 more timeline events
All Facts (17)
COGEXT offers a free tier, its official website is cogextai.com, and it can be installed via pip install cogext.
60%UnverifiedThe author published an audit report on 120 agent output samples as the empirical basis for the precision-first strategy.
50%UnverifiedFor an accountability system, the cost of incorrectly tracking a non-existent commitment is higher than missing an ambiguous one.
50%UnverifiedCOGEXT is a project that establishes an accountability layer for AI agents, used to track commitments made by agents.
50%UnverifiedCOGEXT tracks the lifecycle of each commitment through a state machine: open→due→overdue→fulfilled.
50%UnverifiedCOGEXT's Contradiction Radar is used to detect self-contradictions between an agent's earlier and later statements.
50%UnverifiedCOGEXT is a verification layer tool for authenticating AI Agent behavior, working by extracting commitments and independently verifying them in external systems.
50%UnverifiedThe author states that COGEXT v2.0 passed 8/8 end-to-end (E2E) tests in a production environment.
50%UnverifiedWhat COGEXT can verify depends on the external systems it is connected to, and verifying actions that lack clear external records remains a challenge.
50%UnverifiedThe paradigm offered by COGEXT is to shift verification responsibility from the model itself to independent external systems.
50%UnverifiedCOGEXT's Public Audit Receipts generate a traceable audit record for each verification.
50%UnverifiedCOGEXT's core process consists of two steps: first extracting commitments from Agent outputs, then placing those commitments in external systems for independent verification.
50%UnverifiedCOGEXT's verification targets include Gmail, GitHub, and Webhooks.
50%UnverifiedIf no corresponding evidence is found in external systems, COGEXT blocks the Agent from marking the task as completed.
50%UnverifiedCOGEXT's Kill Switch implements human intervention through Slack's approve/reject interaction.
50%UnverifiedCOGEXT v2.0 contains five core capabilities: a verification engine, a kill switch, a contradiction radar, a failure predictor, and public audit receipts.
50%UnverifiedCOGEXT was shared and published by a developer on Reddit.
50%