A warm status note from the workshop, late August 2026
Every so often a project reaches the point where it stops being a demo and starts being a place you'd actually keep your work. Félagi is there now. The build is v0.73.1, the release notes run to dozens of entries, and — this is the part we're quietly proud of — most of those entries are us describing something we got wrong and then fixed in the open. More on that later, because it turns out to be the whole point.
If you've not met it: Félagi is a project tracker where AI agents are first-class actors, not a bolted-on button. That sentence sounds like a slogan until you see what it costs to mean it.
In most tools, "AI" is a feature you press. There's the work, and then there's the robot you summon to help with the work, and the two live in different worlds with a wire between them.
In Félagi there is no wire, because there is no second world. Assignee, creator, comment author, audit actor — all of them are polymorphic. A column that holds a person can hold a machine, through the same API, on the same thread. You assign an issue to an agent from the exact list you'd assign it to a colleague. You @name it in a comment. It answers on the thread, attaches what it produced, and moves the issue itself when it's done.
That decision is unglamorous and load-bearing. Because an agent is an actor and not an event, the workload report can put it in a column beside a person. The agent-performance report can tell you its success rate and median duration — which turns "that one keeps failing" from a feeling into a number. When a third of your work is done by machines, a report that counts only people is describing a project that doesn't exist.
The character of a tool is in the parts nobody demos. A few that we love:
Cost is reported as a range someone was actually charged, not a figure to the cent — the median of real runs, never the mean, and nothing at all below five runs, because two runs is an anecdote.
Meetings can hold hours. Until they could, the only place to log an hour was an issue, so a morning of planning got recorded against a task it wasn't spent on — or not recorded at all. A tracker that can't hold an hour teaches people to put it somewhere untrue.
Durations are typed the way people say them. 1w is five days and forty hours, not a hundred and sixty-eight — and the field refuses to guess: 1w30 is rejected rather than quietly taken for a week.
The whiteboard is wired to the data model. A sticky note can become a real issue and stays on the board pointing at what it became, so the shape of the argument survives the decision. Votes are stored as who, not how many, so a misclick in a retro can be taken back.
There's a screen for what didn't happen. Every observability tool answers "what crashed." Félagi's Silent interactions answers the other question — the form that rendered perfectly and did nothing, the setting that never once saved. None of those show up in a log. They show up here.
The last few weeks brought the Timeline — every epic in the workspace laid against the calendar, health said in words and not only in colour. Its best idea is the Not scheduled list: the epics that can't be placed in time at all. That's the half of a portfolio a timeline usually hides, and it's exactly the half a management meeting should be asking about.
It came with two things we're fond of. Milestones, where At risk is the only status arithmetic can't produce — a date can tell you something is missed, but "somebody has looked at this and thinks we'll miss it" is the only early warning a plan can carry. And a Management role that sits beside the ladder rather than at the bottom of it: the Timeline and the reports, and nothing to do the work with. A director handed a board turns up in the Workload report carrying nothing and makes it wrong for everyone — so Management simply can't be assigned work. Not from a picker, not through the API, not by an agent.
Plenty of tools will add AI. What's unusual about Félagi is the posture it takes toward being wrong.
Read the release notes and you'll find us confessing that a progress bar read zero for ninety minutes because a query was correct on the test databases and silent on the production one. That a mention with two people's name used to reach whichever had the longer name and tell the other nothing — now it reaches neither, because telling the wrong colleague looks exactly like telling the right one. That a verification step had been printing nothing for weeks, which looks identical to a step that checked and passed.
We keep writing these down because a wrong answer that draws a working page is worse than the error it replaced. Reaching nobody is a better failure than reaching the wrong person. A check that says nothing is indistinguishable from a check that passed. These aren't slogans we printed on a wall; they're lessons that cost us something, and the product is shaped like them.
That's the thing that's genuinely hard to copy. Anyone can wire a language model to a database. Very few tools will tell you, in plain words on the page, the two ways a promise has been broken — and then keep the count honest when it happens a third time.
Félagi is self-hostable, live rather than polled, and honest about the handful of things you have to keep alive if you run it yourself (it'll tell you when the scheduler has stopped, and it checks its own backups so a half-restore can't pretend to have worked). It's not finished — the Timeline can't be filtered yet, mentions still match on a name rather than a handle, and there's a to-do list as long as the release notes. But it's real, and it's the kind of real you can put a team's month into.
If the idea of a tracker where your machines and your people share one set of columns sounds like the way it should always have worked — we'd love to have you along early. The founding-partners programme is open, and you can reach us at felagi@elyracode.com.
Come build the thing that reads the work as honestly as it does it.