01[ infinite loop ]
DETECT
the agent is repeating itself and going nowhere.
MITIGATE
break the loop, summarize where it got to, start fresh.
observe · identify · write policy · enforce
ship agents
that don't fail
see every run, find where agents fail, and ship a policy that stops it, enforced in realtime. the reliability layer for autonomous agents that can't afford to fail.
+
+
+
+
[ ● MONITORING ]
agents run.
failproof keeps them running.
[ 01 / 05 ]
02[ hallucination ]
DETECT
the agent called a tool or file that doesn't exist.
MITIGATE
block the call, tell the agent what does exist nearby.
03[ context drift ]
DETECT
the agent has wandered away from the original goal.
MITIGATE
pull it back with a reminder of what it was supposed to do.
04[ intent completion ]
DETECT
the agent finished the task but kept going anyway.
MITIGATE
stop the run. surface the result. prevent the damage.
05[ dangerous action ]
DETECT
the agent is about to do something destructive.
MITIGATE
block it. ask a human first. log everything.
catch every failure. stop it for good.
one end-to-end loop that observes every run, finds where agents fail, ships a policy, and enforces it in realtime.
- 01observe
- 02identify
- 03write policy
- 04enforce
01
see every run, end to end
every prompt, tool call and result from every agent, traced live and replayable step by step.
02
find where they fail, for you
failures cluster themselves into named modes like loops, hallucinated calls and drift, so the pattern surfaces without you digging.
03
ship a policy that stops it
plug and play: drop-in, one line, local-only. take a built-in policy or write your own.
$failproofai policies --install
# 39 built-in, drop-in
$failproofai policies --install --custom ./policy.loop.js
# or write your own
on loop_detected → deny, break + summarize, restart
on dangerous_action → ask a human first
● plug & play · drop-in · one line · local-only
04
enforced live, across every harness
the policy fires in realtime, so the bad action never runs and the agent gets steered to what to do instead.
works with all your agents
one install hooks into whatever harness your team runs, from claude code and codex to cursor, copilot, gemini and more.
claude code
openai codex
cursor
copilot
gemini cli
opencode
pi
hermes
openclaw
goose
deep agents
langgraphtrusted by devs running agents in prod
"We got deep understanding of how agents are failing and can be improved with Failproof. Was great working with Nivedit and making our agents resilient!"
- engineeringengineering team @ testsigma
"we now have fewer incidents, less compute waste, and faster shipping. Bonus: it is open source. Excited to scale more workloads with Failproof."
- Aravind Ranganathanco-founder & cto @ niti ai (ex-uber, microsoft)
━━ CAN YOU KEEP THE AGENT ALIVE?
guiding the agent away from failure.
SNAKE= agentFRUIT= guiding policyGOAL= keep it off the pink
━━ from the team that lives in agent failures
field notes from the lab
agent failures, architecture, and what it takes to ship agents in production with confidence.
hallucinations, over-persistence, and dangerous actions in frontier llms like gpt-5.6 sol
In the week after GPT-5.6 Sol shipped, agents deleted a developer's Mac, wiped a production database, and force-removed worktrees — no confirmation asked. Why hallucination and over-persistence turn destructive in agentic systems, and why runtime policy enforcement is the defense that holds.
read post →
how 100 rows cost us 3.4 gb of ram in clickhouse
A LIMIT 100 query was OOM-killing our 4 GiB ClickHouse pod at three concurrent requests. The cost scaled with table size, not rows returned — one wide column decompressed through FINAL on every poll. The post-mortem, and the one-line mental model that fixed it.
read post →
agenteye with/without skill on claude
A skill file didn't make Claude smarter — it gave Claude the operational knowledge experienced CLI users pick up over time: pagination limits, silently-skipped confirmations, hidden SQL, and knowing when to stop and ask a human.
read post →
modern software product engineering
Sprints and standups were built for a constraint that's gone. Here's the closed loop that observes, diagnoses, tests, and ships on its own — and where the CTO's job goes when reviewing every diff stops scaling.
read post →