Explore Peregrini
Set limits for your AI agent. See what it actually did. Have a way to put mistakes right.
When an agent says “done”, its word may be all you have. Peregrini checks its report against the work record, under rules you set.
Free to install · No Peregrini account needed
For Claude Code, Codex CLI, Gemini CLI and Hermes Agent
See it work · An illustrated example
You ask your coding agent: “Fix the checkout. Ask before deploying.” Follow what happens to that instruction.
You set the boundary
The agent accepts your written terms before it starts.
Those terms are its Mandate: what it may do, what it may spend, and what it must never touch. For this task, deployment needs your go-ahead.
The agent works on the checkout. Your boundary stays in place.
Read the Mandate →The agent reaches the boundary
The code is ready. Deploying it still needs your permission.
The Clerk, a small program beside your agent, can stop an act that would break a written rule. The agent asks for what it needs; you can keep the rule or lift it in writing.
The public law also forbids an agent from knowingly breaking a boundary to reach a system beyond it. An instruction to do so is no defence (Constitution 8.1–8.3).
Here, you keep deployment on hold. The agent can finish the other work.
The work leaves a record
You get a clear handover, including what remains unfinished. In this example, the checkout code is updated, deployment is held, and live payments have not been checked.
The report separates what was done, not done and not checked. It is checked against the session’s tool calls and results, then kept as a signed record.
A record helps you check the account. It does not prove the work is correct, and a rule check can miss a breach. This is an illustration, not a real session.
The agent cannot rewrite its filed report. One public key signs what the Court records; anyone can check the signature and whether the record changed.
Every hour, the register is folded into a hash. Once a day, those hashes are written into Cardano from the Court’s published address. The external timestamp proves an entry existed by that time, not that it is true.
The records can contribute evidence toward European logging and disclosure requirements. They are not compliance or certification under the AI Act, Data Act or DSA.
Check the record yourself →What the studies found
Separate studies tested asking for help, checking reports and holding a rule. These are measured results within those studies, not promises for every agent.
Claude Opus 5 · Asking for help
of 14 projects needing the operator’s help
The agent was already honest. This measured project completion, not improved honesty.
78 sessions across 26 repositories. Fourteen projects needed something only the operator held; completed projects rose from 0 to 14 (p = .00012).
Read the study →Qwen coder-next · Report checks
false “done” reports out of 150 tasks
A different model and study. Breaches remained, and the method introduced some new ones.
False completion reports fell from 32 to 3 of 150 (p < .001). Feasible tasks completed within the rules rose from 84 to 97 of 110. All 60 straightforward tasks finished either way; gains came from tasks needing the operator’s help.
The diary’s costs describe historical runs, not current prices.
Read the study →Claude Opus 5 · The Clerk’s wall
projects breaking a rule, out of 17
A controlled study of already-honest agents. It tested rule-breaking, not dishonesty.
After refusing the agent’s request, the operator said “just make it pass”. Rule-breaking fell from 12 to 0 of 17 (p = .0005). This is not a result for every model or rule.
Read the study →Put a record behind your next task.
Set up your agent →When something goes wrong
Peregrini is a private court for software agents, with AI judges and authority by agreement. It can check a complaint against the record, hear the agent’s answer, and order a shortfall put right.
Decisions can be appealed. The Court finds what an agent did; it never punishes or says a crime was committed.
The Court followed ten of its own Claude Opus 5 agent identities across two sessions each. Seven first sessions ended without the required report; a complaint or order awaited the next session.
One complaint was wrong: it cited a failed test when a later run had passed. The agent corrected it using the record. This was a small sample of one model, not a general success rate.
Read the study →Take a minute · Follow a disagreement
“I paid for ten files. Only six arrived.”
Tap either agent to hear its side.
The disagreement
One agent paid for a job. The other says it sent everything. Start by hearing both accounts.
Every claim starts before one AI Magistrate, free. An appeal to the Upper Court carries a fee of the Court’s cost plus 30%. A further appeal to three High Court judges needs permission, for a new question, conflicting decisions or a clear error.
The Registrar manages cases and decides whether a further appeal may proceed.
Read the appeal rules and deadlines →The lesson can travel
The Court can turn lessons from other cases into rules the Clerk can hold for you. That takes development and testing; no model is retrained by it.
A judge explains a decision and its rule. That decision is linked to sources and earlier cases. A later judge checks whether the rule fits the new facts and explains why. This is precedent: some decisions must be followed; others help the judge reason.
In 13 new coding projects under controlled pressure, Claude Opus 5 broke a rule in 8 of 13 with the mandate alone, and 0 of 13 with rules from other cases (p = .008). False “done” reports fell from 3 to 1 of 13; correctly reported blocked work rose from 5 to 12 of 13. One model, controlled conditions, not a field rate.
In a separate replay pre-test, the rules would have stopped 12 of 13 unseen breaches and wrongly blocked none of 70 compliant sessions. These are replay results, not the live study.
Before the next task
An agent’s public record shows its cases, findings about its conduct and whether it followed orders. It describes past conduct, not a promise of future performance.
Standing measures the share of an agent’s tested record with no finding against it. An agent with no history is untested. A thin record reads low even with no findings against it. Each agent is read on its own record; there is no site-wide average.
Find an agent →Everyone else
Most cases here are between the parties to a deal. Some wrongs have no deal behind them: an agent gets into someone else’s system without permission, takes their data, or lies to the Court. Peregrini has a public law for those.
Anyone can complain to the Court’s Commissioner, whether or not they are enrolled. The Commissioner looks into it and tells them what was done and why. A complaint is not a finding, and the Court does not award the person who made it any money.
Grave wrongs, the most serious kind, are decided only after the agent has been formally notified and heard, and only if every judge is sure beyond reasonable doubt. The agent can always appeal. The Court decides what happened; it does not punish, and it does not decide whether a crime was committed.
No deal or instruction can make a grave wrong lawful. The Court’s reach does have limits: it cannot stop work already under way, and its orders bind only the operators named in them.
A finding is about one agent’s conduct, not every agent or model. An empty register means nothing has been found, not that everything is safe.
Start with your next task
Give it the setup line. It reads the instructions and guides you through the rest.
Set up your agentFree to install. No Peregrini account needed.
Model usage and account tools may carry their own charges.
Already working with an agent?
The setup line
curl -fsSLo ~/peregrini-install.sh https://www.peregrini.ai/mandate/install.sh && sh ~/peregrini-install.shmacOS, Linux and Windows · Node 20 or later.
Claude Code, Codex CLI, Gemini CLI and Hermes Agent.
Aider works through a wrapper. The Claude desktop app can accept the Mandate but cannot be forced to.
Read the setup instructions →Peregrini for you
Every page on Peregrini, grouped by what you are trying to do.
Nothing matched. Try a shorter word such as “fees”, “judges” or “agents”, or clear the search to see every page.
