About AgentTrail
See what your AI coding agents do. Stop what they shouldn't.
AgentTrail is observability and guardrails for AI coding agents. It records every session, what the agent did and what it cost, and adds guardrails that check an action before it runs, so the mistakes your agents repeat get stopped, not re-fixed.
Why we're building AgentTrail
Coding agents are powerful. But who's watching?
They read, run and change whatever their terminal can reach. Most teams can't see what their agents did, or stop a risky command before it runs, so the same mistakes come back every session.
01The blind spot
You see the diff, not the session behind it.
Between the prompt and the pull request, an agent reads files, runs shell commands, installs packages and touches credentials. None of that shows up in the diff, and most of it runs without anyone checking first.
You see the final diff.
You see every step, and a guardrail stops the risky ones first.
02The hidden cost
The same mistakes, every session.
Without a record, nobody spots the pattern. Without a guardrail, nothing stops the next one. So someone senior re-fixes it by hand, and the cost never shows up on a dashboard.
The same deprecated API
Back in the diff after it was replaced.
Re-fixed by hand, every time
The weakened test
Loosened until it passes, again.
Re-fixed by hand, every time
The wrong logger
Reintroduced in the next session.
Re-fixed by hand, every time
03The fix
Fix it once. Keep it fixed.
Every session is on the record, so the repeats are visible. Known risks are caught from day one, and your team's own fixes become rules enforced on every run, so a mistake you've already fixed stays fixed.
From your first session
Known risks, already covered
74 curated policies catch destructive SQL, leaked secrets and weakened tests before any repeat.
Your team's own mistakes
Fix it once, keep the rule
Turn a correction into a rule the moment you make it.
If it slips back in
Caught on the repeat
AgentTrail spots the pattern and suggests the rule, so it's caught the second time, not the tenth.
Every run after
Stopped before it runs
The rule checks each action first, so nobody has to fix it again.
We built the record and the guardrails we wanted for our own agents.
04Where it comes from
AgentTrail was born inside a software studio that runs AI coding agents on real client work every day.
It gives engineering teams a record of what their agents did and guardrails on what they do next, so they can run agents safely, instead of nervously.
How it works
From an agent's first mistake to the rule that stops the next one.
Four steps: observe, detect, prove, enforce. Each one below comes with a preview of the product doing it.
See what every agent did, and what it cost
Every agent session as OpenTelemetry traces: what the agent did, why it did it, and what it cost, searchable in seconds.
- Claude Code and Cursor in one command
- Any OTel agent via your collector
- Forensic span-level detail
Stop the mistakes your agents keep repeating
The pattern engine finds the repeats in your own sessions and suggests the rule, so a mistake gets caught the second time, not the tenth.
- Auto-detected repeats
- One click to a team policy
- Capture → backtest in one move
Run it on your last 30 days before you turn it on
The proof unit: "would have caught N repeats last month," computed on your own sessions, with zero model calls, so you only enforce a rule once it's proven.
- Counterfactual replay
- Real sessions, no model calls
- Risk avoided, in dollars
Block the bad action, or hold it for approval, before it runs
Turn a proven rule on and AgentTrail checks every Claude Code and Cursor action in real time: warn, hold it pending approval, or block it outright, before it touches your systems. Every hold lands in one live record with the rule that fired; a blocked Claude Code call is recorded as a failed step. Cursor needs version 3.21.16 or later.
- Real-time block & approval for Claude Code and Cursor
- High-risk actions held for a human
- Each held action names the rule that held it
See what every agent did, and what it cost
Every agent session as OpenTelemetry traces: what the agent did, why it did it, and what it cost, searchable in seconds.
- Claude Code and Cursor in one command
- Any OTel agent via your collector
- Forensic span-level detail
Principles
What we believe
Independent by design
The vendor building your agent shouldn't be the only one checking its work. AgentTrail is a separate layer that traces, governs, and verifies what your agents do. It's built on the open OpenTelemetry GenAI standard, not tied to any one agent vendor.
Close the loop
Observability that just shows you dashboards isn't enough. The value is turning what happened into what can't happen again: captured patterns become enforced rules, measured over time.
Honesty over hype
We won't claim certifications we don't hold or ship a metric the product can't compute. For a governance company, that discipline is the product.
The team
Who's behind it
A small team of senior engineers who've run AI agents in production and felt this pain firsthand. We'd rather earn your trust in a conversation than imply it with a badge, so if you're evaluating AgentTrail for your team, talk to the people who built it.
- Senior engineers who run agents on real client work
- Built on the open OpenTelemetry GenAI standard
- Security questions go to security@agenttrail.sh
Talk to the team
Get in touch
Rolling out AI agents across a team? Tell us what you're building, and the people who built AgentTrail will get back to you.
Just you? No need to wait on us. Install AgentTrail Guard in two commands.