Who Decided That?

A booking system starts telling patients "sorry, that slot is no longer available" at the last step of booking. The cause turns out to be a 60-second cache of each physio's availability, which keeps offering slots that were just booked through another server. The cache was a sensible fix for a slow page. Nobody remembers asking for it.

That system is Harbour, the sample project behind Tracker, and the cache was an agent's idea. It wrote down the decision, why it made it, and that it was high impact. Most of the time, nobody writes anything down at all.

Coding agents now write a large share of the code: on average, around 47% of professional developers' work code is fully written by agents, according to JetBrains' 2026 research. And they don't just type. They choose caches, retention rules, session lengths, database upgrades: dozens of design calls nobody asked for.

Git records what changed and who pushed it, not why, or whether a human ever asked for it. The agent's chat records what was said, until the session closes. Jira records what humans planned, not what agents decided while building it. ADRs record the decisions someone bothered to write down.

So when a client, an auditor or your own CTO asks "who decided that, and with what context?", the honest answer is usually "the agent, probably".

Tracker rests on one split. You state intent — goals and constraints in your own words, kept verbatim. Agents decide how to build it, and when they fill a gap, they say so.

Every decision carries its origin:

• ● Human: you said it, in an intent or an answer.
• ◐ Delegated: you were asked and said "you decide".
• ○ Inferred: the agent derived it from your words, with a pointer to them.
• ◆ Autonomous: the agent decided on its own.

The one number that matters most is drift: autonomous decisions nobody has reviewed that are high impact or one-way. That's where your system quietly diverges from anything a person asked for.

Tracker keeps one linked record from your words, through the questions agents ask and the decisions they make, to the commits and releases GitHub reports. Agents connect over MCP, with plugins for Claude Code, Cursor, Codex, Gemini CLI, VS Code Copilot and omp; the questions and the risky decisions wait in your Inbox. Every write lands in a hash-chained ledger with public signed checkpoints, so anyone can verify the record without trusting us.

You can see it working, with no sign-in, in Harbour: a clinic booking system built by agents from two people's intent. Of its 42 decisions in force, 14 were the agents' alone, and four of those are drifting. When patient phone numbers turned up in the API logs, Tracker traced it in seven steps: from the change that caused it, through the earlier fix it was part of, to the human decision and the original request behind that — and forward to the fix.

Tracker can't prove everything, and it says so. Prompts captured by a tool on your machine are marked unverified, because an agent on that machine could read the same token. An agent named only in a commit trailer is claimed, not verified; agents with their own credential or a registered signing key are verified. And the ledger began on 27 October 2026: anything before a project's first entry is a snapshot the ledger can show but not prove.

It also doesn't block your agents. It records and surfaces; the decisions about what to do next stay with you.

We're looking for a small number of teams building with agents — agencies building for clients, and regulated teams who will be asked how AI-written change is reviewed — to put one real project into Tracker for 90 days, with a weekly 30-minute Inbox review together.

Read more about Tracker, or tell us which project you'd put in first.