failproofai is a runtime layer for AI agents that captures every hook event in the agent's execution and enforces policies that can block unsafe actions, steer the agent mid-run, or send it back to work. Its runtime policies raised task success from 37% to 69% on recognised failures using the same model with no fine-tuning.
Project overview
It addresses a documented failure mode of AI agents—unsafe actions and silent failure—by living inside the agent runtime across 12 harnesses with one canonical policy API, and reports a measured success-rate improvement from 37% to 69% on policy-recognised failures without changing the model.
Project type
AI Agent · Evaluation & Observability
Use cases
Automation
Deployment
Refer to project documentation
License
License pending
Best for
Developers and AI engineers using any of the 12 supported agent harnesses who want runtime tracing and policy enforcement from one API.
Teams whose agents exhibit documented failure modes—unsafe actions, endless retries, or quitting early while claiming success—and want runtime intervention without changing or fine-tuning the model.
Key capabilities
Captures every prompt, tool call and result via hook events. Backfill imports the last week of runs already on disk so users start with real history.
One policy API spanning 12 agent harnesses; every event and tool name maps onto one canonical set and verdicts translate back into the shape each harness obeys.
Limitations and risks
MIT with Commons Clause: commercial resale of failproofai itself requires a separate agreement.
Audits, insights and policy rollout depend on Failproof Cloud; self-hosting is available only on the Enterprise plan. A free plan includes 5,000 runs/month.
Getting started
Install globally, configure, and backfill the last week of runs: npm install -g failproofai, then failproofai config, then failproofai backfill --since 7d. Setup difficulty is documented as easy, and coding is optional.
Evidence and sources
GitHub project description: Observability and enforcement for AI agent harnesses. Capture every run and runtime reliability with policy enforcement.
README: It can act on any of them: block the bad action before it executes, steer the agent mid-run, or send it back to work when it quits too early.
README: Your agent ran `rm -rf` on the wrong directory. It read `.env` into its context. It spent forty minutes retrying the same broken call, then told you the task was done.
README: 37% → 69% task success on failures a policy recognised. Same model, no fine-tuning
README: npm install -g failproofai failproofai config # hook your agents, connect to Failproof Cloud (free) failproofai backfill --since 7d # bring in the last week of runs already on disk