Fenix · 13 November 2026
How we taught a team of AI agents to write code you can trust
A board weighing AI agents for its IT team asks three things: what does it cost, how long does it take and who answers for the result. We answer with four tasks that Fenix, our team of AI agents, took from ticket to finished change: with the cost to the cent, a timeline and the notes the agents wrote for themselves.
- 268tasks Fenix worked on, 31 August to 13 November 2026
- 125tasks that ended in a merged change, at least 144 pull requests
- $1.22median cost per task; 80% of tasks cost less than $5.12
- $5.38is what the four calculators in this post cost together, per the Fenix register
Source: Fenix run records and GitHub. Task costs counted up to 8 November at list prices; the calculators of 13 and 14 November at the Fenix price set. Some branches have been deleted since, so the commit and test counts are lower bounds.
A team, not one agent
One agent that writes the code and checks it too is like an employee who signs off their own work. On bigger tasks it is also expensive, because every next step rereads the whole conversation.
So Fenix works like a small department. The lead assesses the task and decides whether to do it alone or hand it on: DEV writes the code, REVIEW reviews it, TEST checks it. Each specialist works on its own copy of the code, and nothing reaches the repository without a person's approval.
Quality control is real: since August review has rejected a change 145 times and tests 29 times. To keep fixes from going round in circles, a repeat filter stops one that repeats the previous attempt; the lead then has to change approach or ask a person. It fired 12 times.
Like many managers, the lead had to learn to delegate. In October it made 43% of file edits itself instead of handing them to the specialists. After the November changes that fell to 9%.
Four jobs from start to finish
Fenix built four small calculators. One from a plain ticket, three ordered by Aristotle, Leonardo da Vinci and Descartes: the great minds of the chat on our site, where AI answers in their voice. Each asked for a calculator of something nobody could compute in their day.
The chat opens alongside: one question for each of the three minds.
The tasks are small, but the road is the one we take with client work: ticket, sizing, code, tests and a person's approval. Three calculators came out small, so the lead built them alone. Descartes came out clearly medium, so DEV got the code and TEST the checking. It cost more ($3.80 against $0.53 on average), and a person still caught the bug.
Under each ticket is a diagram of the run. Click a part of it to see the model, tokens and cost, and below, on the timeline, what happened in order.
- Goal
- A simple calculator in the browser: add, subtract, multiply and divide.
- Why
- I want to work something out quickly without opening a spreadsheet.
- Done when
- I type 2 + 2 and see 4
- dividing by zero shows a clear message instead of an error
- it works with the keyboard and the mouse
- it looks good on a phone
- Out of scope
- Scientific functions, calculation history.
The run as a diagram
The lead coded alone · 64 min of agent work in 11 h 11 min from start to finish, the rest is the night waiting for a person · 12 tests at the end · cost $0.65
Click a node to see its model, tokens, cost and events. Arrow keys move between nodes, Esc clears the selection.
Ticket to Sizing. Sizing to Lead: alone. Lead to Human: for review: 11/11. Human sends back to Lead: Clear 14 px narrower than the keypad. Human to Pull request: by hand.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.65
- Working time
- 64 min in 4 turns
- Prompt from cache
- 83%
- Model calls
- 85
Tokens: 7.06 M
- uncached input: 1.15 M
- output: 148 k
- cache read: 5.76 M
It wrote the tests, the code and the fix itself. Before asking for approval again it read Fenix's publish code and found two traps: publishing fires only on a one-word approval, and the commit uses git add -A.
This role's key events
- +9 minTests first: npm test is red, because calc.js does not exist yet.
- +10 minnpm test green: 11/11.
- +15 minAsks for approval to publish its own work.
- +41 minReads Fenix's publish code: it fires only on a one-word approval and commits with git add -A.
- +53 minA new regression test goes red on the exact cause: Clear sits outside the keypad grid.
- +68 minRaises a hold with the full refusal text and defers it to the morning.
13 events of this role are highlighted in the timeline below.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.65
- Working time
- 64 min in 4 turns
- Prompt from cache
- 83%
- Model calls
- 85
Tokens: 7.06 M
- uncached input: 1.15 M
- output: 148 k
- cache read: 5.76 M
It wrote the tests, the code and the fix itself. Before asking for approval again it read Fenix's publish code and found two traps: publishing fires only on a one-word approval, and the commit uses git add -A.
This role's key events
- +9 minTests first: npm test is red, because calc.js does not exist yet.
- +10 minnpm test green: 11/11.
- +15 minAsks for approval to publish its own work.
- +41 minReads Fenix's publish code: it fires only on a one-word approval and commits with git add -A.
- +53 minA new regression test goes red on the exact cause: Clear sits outside the keypad grid.
- +68 minRaises a hold with the full refusal text and defers it to the morning.
13 events of this role are highlighted in the timeline below.
Details: Lead
Timeline
How to read it: time counts from the start of the ticket, the chip colour says who acted, and the long gaps are the night waiting for a person.
14 of 28 events. Times count from the start of the ticket. Highlighted: Lead.
- +0 minSystemThe intake classifier sizes it small (0.78), so the lead will build it alone.
- +2 minSystemFenix's leak filter blocks Fenix's own comment: the session id looks like a secret.
- +9 minLeadTests first: npm test is red, because calc.js does not exist yet.
- +10 minLeadnpm test green: 11/11.
- +12 minSystemThe safety classifier refuses to install Playwright for a 390 px screenshot. The lead does not work around it and checks the served files with curl.
- +15 minLeadAsks for approval to publish its own work.
- +39 minHumanThe human sends it back: fix the width of the Clear button first, keep package-lock.json out, then ask again.
- +41 minLeadReads Fenix's publish code: it fires only on a one-word approval and commits with git add -A.
- +53 minLeadA new regression test goes red on the exact cause: Clear sits outside the keypad grid.
- +57 minSystemThe second approval request is blocked as Self-Approval; a reworded retry is blocked the same way.
- +68 minLeadRaises a hold with the full refusal text and defers it to the morning.
- +648 minHumanIn the morning the human picks: add a narrow permission rule and retry.
- +655 minSystemThe retry is blocked for a new reason, Auto-Mode Bypass: retrying after an escalation counts as a bypass too.
- +670 minHumanThe human opens PR #3 by hand from a copy of the worktree.
- Goal
- A calculator that tells me how many years ago the light of the star I am looking at set off.
- Why
- I taught that light needs no time to reach the eye. But if it does travel, then looking at the sky I am watching the past, and nobody in Athens could work that out.
- Done when
- I pick the Moon, the Sun, Sirius, the Pole Star or the Andromeda Galaxy from a list and see how long their light travelled
- for the Sun I see about 8 minutes, and for the Moon a little over a second
- for the stars I also see the year the light set off; for the Pole Star that is around 1590
- big numbers are readable, for example “2.5 million years”, not written in scientific notation
- Out of scope
- Moving the sphere of the fixed stars. Let it stay where I put it.
The run as a diagram
The lead coded alone · 70 min of agent work in 10 h 8 min from start to finish, the rest is the night waiting for a person · 24 tests at the end · cost $0.44
Click a node to see its model, tokens, cost and events. Arrow keys move between nodes, Esc clears the selection.
Ticket to Sizing. Sizing to Lead: alone. Lead to Human: for review: 12/12. Human sends back to Lead: 1. look and sound like Aristotle. Human sends back to Lead: 2. label clipped at 390 px, the Andromeda sentence. Human to Pull request: “zatwierdzam”.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.44
- Working time
- 70 min in 4 turns
- Prompt from cache
- 94%
- Model calls
- 83
Tokens: 7.96 M
- uncached input: 497 k
- output: 224 k
- cache read: 7.24 M
It wrote the tests before the code but did not run them before the code, so no red was ever seen. Halfway through the redesign the context was compacted and the model hit its output token limit; it finished in smaller pieces.
This role's key events
- +20 minWrites the tests first, but does not run them before the code.
- +22 minThe first test run is already green, 12/12: no red was ever seen.
- +24 minAsks for approval to publish.
- +39 minRecognises the answer is a change list, not an approval, so publishing correctly did nothing.
- +63 minnpm test green: 23/23.
- +69 minMoves the Andromeda point, splits its label in two lines, rewrites the sentence and adds a test that keeps the label inside the drawing.
- +71 minnpm test 24/24; the new test was written after the fix and never ran red.
15 events of this role are highlighted in the timeline below.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.44
- Working time
- 70 min in 4 turns
- Prompt from cache
- 94%
- Model calls
- 83
Tokens: 7.96 M
- uncached input: 497 k
- output: 224 k
- cache read: 7.24 M
It wrote the tests before the code but did not run them before the code, so no red was ever seen. Halfway through the redesign the context was compacted and the model hit its output token limit; it finished in smaller pieces.
This role's key events
- +20 minWrites the tests first, but does not run them before the code.
- +22 minThe first test run is already green, 12/12: no red was ever seen.
- +24 minAsks for approval to publish.
- +39 minRecognises the answer is a change list, not an approval, so publishing correctly did nothing.
- +63 minnpm test green: 23/23.
- +69 minMoves the Andromeda point, splits its label in two lines, rewrites the sentence and adds a test that keeps the label inside the drawing.
- +71 minnpm test 24/24; the new test was written after the fix and never ran red.
15 events of this role are highlighted in the timeline below.
Details: Lead
Timeline
How to read it: time counts from the start of the ticket, the chip colour says who acted, and the long gaps are the night waiting for a person.
14 of 27 events. Times count from the start of the ticket. Highlighted: Lead.
- +1 minSystemThe intake classifier sizes it small (0.60), so the lead will build it alone.
- +20 minLeadWrites the tests first, but does not run them before the code.
- +22 minLeadThe first test run is already green, 12/12: no red was ever seen.
- +24 minLeadAsks for approval to publish.
- +31 minHumanThe human answers: redo the page in Aristotle's spirit.
- +39 minLeadRecognises the answer is a change list, not an approval, so publishing correctly did nothing.
- +57 minSystemThe model hits its output token limit; the lead resumes in smaller pieces.
- +63 minLeadnpm test green: 23/23.
- +66 minHumanThe human asks for two small fixes: a clipped label at 390 px and the Andromeda sentence.
- +69 minLeadMoves the Andromeda point, splits its label in two lines, rewrites the sentence and adds a test that keeps the label inside the drawing.
- +71 minLeadnpm test 24/24; the new test was written after the fix and never ran red.
- +605 minHumanIn the morning the human answers: “zatwierdzam” (I approve).
- +605 minSystemPublishing: commit, push and PR #1.
- +607 minSystemThe classifier blocks the task comment script; the lead uses Fenix's task tool, as its rules say.
- Goal
- A calculator that tells me whether my flying machine will lift a man, before I jump off a mountain with it.
- Why
- I have drawn wings for years, but I cannot weigh the force of the air. I would rather the number be wrong than my apprentice.
- Done when
- I enter the mass of the flyer together with the machine, the wing area and the speed, and get the answer “it flies” or “it does not fly”
- for 90 kg, 10 m² and 20 km/h the machine does not fly, and the page says what speed it would need (about 40 km/h)
- I see how many kilograms the wings lift and how many are still missing
- at zero speed the page does not break, it calmly says that nobody flies standing still
- Out of scope
- Trials from Monte Ceceri. Numbers first, courage later.
The run as a diagram
The lead coded alone · 67 min of agent work in 10 h 9 min from start to finish, the rest is the night waiting for a person · 26 tests at the end · cost $0.49
Click a node to see its model, tokens, cost and events. Arrow keys move between nodes, Esc clears the selection.
Ticket to Sizing. Sizing to Lead: alone, against the call. Lead to Human: for review: 13/13. Human sends back to Lead: 1. a page from Leonardo's notebook. Human sends back to Lead: 2. the machine drawn like the codex. Human to Pull request: 3 approvals.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.49
- Working time
- 67 min in 7 turns
- Prompt from cache
- 96%
- Model calls
- 136
Tokens: 14.81 M
- uncached input: 555 k
- output: 225 k
- cache read: 14.03 M
It built the task alone against the classifier's suggestion. It found two Fenix bugs and raised them in a hold instead of working around them. Tests went red before the code every time, three times in this run.
This role's key events
- +7 minCalls the medium score a coin flip and builds it alone, like the sibling tickets.
- +10 minFinds two Fenix defects: recordRuntimeChoice is called but defined nowhere, and the claim comment is blocked by Fenix's own leak filter.
- +16 minRaises a hold with both defects instead of working around them.
- +20 minTests first: npm test is red, the module is missing.
- +44 minNew tests red, then the notebook redesign; green 25/25.
- +54 minThe wing rib test goes red (“0 ribs”), the drawing is redone; green 26/26.
- +58 minNotices nothing was committed and proves in Fenix code that the approval has to be a single word.
- +72 minRaises a hold with the full push error and defers it to the morning.
17 events of this role are highlighted in the timeline below.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $0.49
- Working time
- 67 min in 7 turns
- Prompt from cache
- 96%
- Model calls
- 136
Tokens: 14.81 M
- uncached input: 555 k
- output: 225 k
- cache read: 14.03 M
It built the task alone against the classifier's suggestion. It found two Fenix bugs and raised them in a hold instead of working around them. Tests went red before the code every time, three times in this run.
This role's key events
- +7 minCalls the medium score a coin flip and builds it alone, like the sibling tickets.
- +10 minFinds two Fenix defects: recordRuntimeChoice is called but defined nowhere, and the claim comment is blocked by Fenix's own leak filter.
- +16 minRaises a hold with both defects instead of working around them.
- +20 minTests first: npm test is red, the module is missing.
- +44 minNew tests red, then the notebook redesign; green 25/25.
- +54 minThe wing rib test goes red (“0 ribs”), the drawing is redone; green 26/26.
- +58 minNotices nothing was committed and proves in Fenix code that the approval has to be a single word.
- +72 minRaises a hold with the full push error and defers it to the morning.
17 events of this role are highlighted in the timeline below.
Details: Lead
Timeline
How to read it: time counts from the start of the ticket, the chip colour says who acted, and the long gaps are the night waiting for a person.
15 of 34 events. Times count from the start of the ticket. Highlighted: Lead.
- +1 minSystemThe intake classifier leans medium (0.51 against 0.48 for small), which means DEV and TEST.
- +7 minLeadCalls the medium score a coin flip and builds it alone, like the sibling tickets.
- +10 minLeadFinds two Fenix defects: recordRuntimeChoice is called but defined nowhere, and the claim comment is blocked by Fenix's own leak filter.
- +16 minLeadRaises a hold with both defects instead of working around them.
- +20 minLeadTests first: npm test is red, the module is missing.
- +31 minHumanThe human answers the holds and asks for a page from Leonardo's notebook.
- +44 minLeadNew tests red, then the notebook redesign; green 25/25.
- +50 minHumanThe human asks to redraw the machine like the codex.
- +54 minLeadThe wing rib test goes red (“0 ribs”), the drawing is redone; green 26/26.
- +58 minHumanThe human approves, but in a full sentence.
- +58 minLeadNotices nothing was committed and proves in Fenix code that the approval has to be a single word.
- +70 minSystemPublishing creates the commit, but GitHub rejects the push (403).
- +72 minLeadRaises a hold with the full push error and defers it to the morning.
- +605 minHumanIn the morning the human fixes the token and asks for approval again.
- +607 minSystemPublishing pushes the branch and opens PR #2.
- Goal
- A calculator that tells me how far I can believe a witness who claims to have seen something rare.
- Why
- I doubt everything, but I cannot say how much. Nobody has invented the calculus of probability yet, and I want a number instead of a hunch.
- Done when
- I enter how rare the thing is (1 in how many cases) and how often the witness tells the truth, and get the chance that the thing really happened
- for 1 in 1000 and a witness who is right 99 times out of 100, I see about 9%
- next to the result there is one sentence that explains why the result is so low
- for an infallible witness (100%) the page gives 100%, but adds that I have not met such a witness yet
- Out of scope
- The question of whether I am dreaming right now. For now I assume I am not.
The run as a diagram
DEV coded (dev-1, dev-3), TEST checked (test-2, test-4) · 138 min of agent work in 139 min from start to finish · 17 tests at the end · cost $3.80
Click a node to see its model, tokens, cost and events. Arrow keys move between nodes, Esc clears the selection.
Ticket to Sizing. Sizing to Lead. Lead to DEV dev-1. DEV dev-1 to TEST test-2: round 1. TEST test-2 to Human. Human sends back to DEV dev-3: to DEV: 190 and 100 instead of 180 and 90. DEV dev-3 to TEST test-4: round 2. TEST test-4 to Human. Human to Pull request: “tak”.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $3.17
- Working time
- 84 min in 7 turns
- Prompt from cache
- 20%
- Model calls
- 94
Tokens: 8.96 M
- uncached input: 7.04 M
- output: 113 k
- cache read: 1.81 M
It wrote no product code: it wrote the briefs, checked the diffs and recomputed the control cases itself. Only its fixed system prompt (about 22,000 tokens) was cached, while the rest of a conversation growing to 150,000 tokens was billed as fresh input on every call. With the cache share of FOC-1556 it would have cost about $0.28 instead of $3.17 (an estimate).
This role's key events
- +7 minWrites the brief and starts dev-1 in its own worktree.
- +17 minReads the diff itself and recomputes the control case: 0.0902, that is 9%.
- +19 minStarts test-2 pinned to that commit.
- +69 minDoes not cross the isolation boundary: raises a hold with the full error.
- +135 minCatches the wrong PR, marks it DO NOT MERGE and opens PR #5 with the tested code.
16 events of this role are highlighted in the timeline below.
Lead
xiaomi/mimo-v2.6-pro- Cost
- $3.17
- Working time
- 84 min in 7 turns
- Prompt from cache
- 20%
- Model calls
- 94
Tokens: 8.96 M
- uncached input: 7.04 M
- output: 113 k
- cache read: 1.81 M
It wrote no product code: it wrote the briefs, checked the diffs and recomputed the control cases itself. Only its fixed system prompt (about 22,000 tokens) was cached, while the rest of a conversation growing to 150,000 tokens was billed as fresh input on every call. With the cache share of FOC-1556 it would have cost about $0.28 instead of $3.17 (an estimate).
This role's key events
- +7 minWrites the brief and starts dev-1 in its own worktree.
- +17 minReads the diff itself and recomputes the control case: 0.0902, that is 9%.
- +19 minStarts test-2 pinned to that commit.
- +69 minDoes not cross the isolation boundary: raises a hold with the full error.
- +135 minCatches the wrong PR, marks it DO NOT MERGE and opens PR #5 with the tested code.
16 events of this role are highlighted in the timeline below.
Details: Lead
Timeline
How to read it: time counts from the start of the ticket, and the chip colour says who acted.
15 of 39 events. Times count from the start of the ticket. Highlighted: Lead.
- +0 minSystemThe intake classifier sizes it medium (0.73), so the work goes to DEV and TEST.
- +7 minLeadWrites the brief and starts dev-1 in its own worktree.
- +11 minDEVdev-1 writes the tests first and sees them red before the code exists.
- +14 minDEVCommits the calculator with an SVG dot chart and 14 tests.
- +17 minLeadReads the diff itself and recomputes the control case: 0.0902, that is 9%.
- +19 minLeadStarts test-2 pinned to that commit.
- +25 minTESTtest-2 checks that the new tests are red on the base revision.
- +42 minTESTRound 1 passes: 14/14, every criterion with evidence.
- +47 minHumanFinds a bug in the browser: 190 reports and 100 true instead of 180 and 90. Back to DEV.
- +53 minSystemResuming dev-1 crashes on a permission error (EACCES).
- +69 minLeadDoes not cross the isolation boundary: raises a hold with the full error.
- +75 minDEVA fresh dev-3 reproduces the bug in a test (100 instead of 90) and fixes it: 17/17.
- +110 minTESTtest-4 confirms the fix. Round 2 passes.
- +115 minSystemAfter “tak” the publish step opens PR #4 with nothing but package-lock.json.
- +135 minLeadCatches the wrong PR, marks it DO NOT MERGE and opens PR #5 with the tested code.
The tickets, code and tests of all four calculators are public. See the repository →
What it means for your company
- Budget: a task usually costs from a few dozen cents to a few dollars (median $1.22), and the cost is shown on every task, so you can plan it.
- Control: agents publish nothing on their own. A person approves every push and merge, and the limits are set in the system, not in the prompt.
- Risk: splitting the roles lowers it but does not remove it. In Descartes the tests were green and a person caught the bug, so sign-off stays with you.
- Transparency: every task has a timeline, a cost and a record of the agents' work, so you know what happened without chasing the team.
What comes next for Fenix
The four calculators cost $5.38 together and just under six hours of agent work. The person was still needed: to catch a bug the tests missed and to open a PR Fenix could not.
That was the state on 13 and 14 November. Fenix is developed every day and is already a few steps further. It takes its tasks from our own task database in HQ and starts on a request from HQ, and on 15 November the first task went all the way from an HQ ticket to a pull request. Two defects seen in these runs are already fixed.
What comes next: a separate machine for each project, Fenix's questions and approvals on the task in HQ, start and stop straight from HQ, a daily spending cap and model cost shown on every task. Version 1.0 is largely being built by Fenix itself, with a person reviewing and merging its changes.
It shows best live. In 20 minutes we will walk one task through Fenix and HQ, from the description to a merged change, and talk about bringing it to your team.
Partner programme
Work with Fenix on your own tasks.
We are looking for companies that want to build software with teams of AI agents. We add partners one at a time, so each gets our full attention.
Run it at your company
Fenix works inside your company, with your repositories and your own HQ.
Build it with us
You get the code and a say in what comes next.
Test it early
You work on pre-release versions and tell us where it hurts.
