Baseline and target
Every goal starts from a measured baseline, not from a feeling that things are bad. Without a baseline there is nothing to compare the result against, so the Factory refuses to start.
Stigmera Factory
The loop that owns the outcome
Stigmera Factory takes a goal — not a prompt — and runs the whole loop: hypotheses → tickets → code → gates → deploy → measure → next iteration. This page is the detail behind that sentence.
01 · Goals
A goal is a metric, not a wish
You do not describe a feature. You pick a metric from a fixed catalogue, state where it stands today, where it should get to, and what must not get worse on the way.
An antagonist metric
Each metric is paired with one that would suffer if it were gamed. Sign-ups up while retention collapses is a failed hypothesis, not a win. This is what keeps Goodhart’s law out of the loop.
The measurement is tested first
Before a goal runs, the measurement is checked against itself: split identical traffic in two and confirm the instrument reports no difference. An instrument that finds an effect where there is none cannot referee anything.
02 · Hypotheses
The agent proposes. The platform rules.
A hypothesis moves through proposed → running → verdict. The last step is the one that matters, and it is deliberately taken away from the agent that authored the work.
- An agent cannot mark its own work confirmed. That verdict is reserved for the platform, which issues it from the measurement.
- A verdict cites numbers. Not a summary of what the agent believes happened.
- A negative result is a result. Rejected hypotheses stay on the record — that is how the next one gets better.
- The verifier is protected. Work that would weaken the thing doing the checking is caught before it lands.
03 · Delivery
Under the loop sits a full SDLC
The interesting part of autonomy is not writing code — it is everything that has to be true before code is allowed to reach production. Tickets, branches, agent sessions, gates, deploy windows, incident handling and a retry policy that depends on why something failed. As of today that pipeline runs 76 steps across eight kinds.
Standards with teeth
Engineering standards are enforced by tests in the pipeline, not by a document nobody reads. Work that violates one comes back to the ticket.
It deploys itself
The platform ships its own updates through its own pipeline, blue-green, in cluster. Everything on this site went through that same line.
Every step is readable
A ticket carries its own history: who took it, which gate failed, how long each stage took, what the deploy reported. Nothing about a run is folded away.
04 · Economics
Cost per delivered ticket, out in the open
Autonomy that costs more than it delivers is a demo. The Factory prices its own work: what a delivered ticket cost, which runs burned budget without shipping anything, and where the money went.
- Cost per delivered ticket is a first-class metric. It is tracked over time like any other, and it is expected to fall.
- De-agentisation is the lever. When a step turns out to be mechanical, it moves out of the model and into plain code — cheaper, faster, deterministic.
- Goals carry budget caps. A goal that runs out of budget stops and reports; it does not quietly keep spending.
- Throughput is paced. Work is scheduled against real capacity windows rather than launched all at once and left to thrash.
05 · Memory
The factory gets better at being a factory
Agents coordinate through traces they leave in a shared environment — tickets, branches, standards, measurements, notes. That environment is also where the system keeps what it has learned.
A knowledge bank
What worked, what failed and why is written down where the next session will read it — instead of being re-derived from scratch every run.
Consolidation
Notes are periodically merged and pruned. Memory that only grows becomes noise, and noise is worse than nothing.
Benchmarks with thresholds
Sessions are scored and the harness is benchmarked against thresholds, so a change that makes the factory worse is visible as a regression.
06 · Controlled autonomy
A ladder, not a switch
Autonomy is granted in steps and can be taken back. Today the delivery loop runs autonomously and deploys wait for a human. Closing the loop end-to-end on business goals is in progress — and we ship it the same way we ship everything else: in public.
- Deploys are on a manual trigger. By design, not by omission.
- Autonomy is scoped by mandate. A level is granted for a scope and revoked when the evidence stops supporting it.
- Budget caps bound every goal. Spending is a guardrail like any other.
- Full goal autonomy: in progress. Stated plainly, because the alternative is a claim we cannot back.
FAQ
The questions we actually get
How is this different from Devin, Copilot or Cursor?
They sell an engineer: a task goes in, a pull request comes out, and a human decides what happens next. We sell the loop: a goal goes in, and hypotheses, code, deploys and measurements keep coming until the metric moves. Different unit of work — theirs is a task, ours is a metric.
Is this Lovable or Bolt?
App builders hand you a project and stop. The Factory keeps the product and keeps working on it — measuring, iterating, shipping the next version.
Is this like factory.ai?
factory.ai builds an autonomy stack for enterprise engineering teams. We build a product factory for solo builders and small teams. Different user, different unit of work.
Isn’t this just AutoGPT again?
AutoGPT had no measurement and no foundation under it. Here the only source of truth is the metric — each one paired with an antagonist metric so the loop cannot game itself — verdicts are issued by the platform from the numbers, and underneath sits a full SDLC with enforced gates.
Autonomous agents sound dangerous. Are they?
It is the right question, and the answer is mechanisms rather than reassurance. Deploys stay on a manual trigger. Standards are enforced by tests. Every metric has an antagonist metric. An agent can propose a verdict but never confirm one — the platform does that from the numbers. Goals carry budget caps. Every step is a ticket you can read.
Waitlist
Put a goal in front of it
The Factory is in private access while the goal loop closes. Join the waitlist and we will write when it opens up.
We keep your email, plus the country and network address the request came from, so we know where interest comes from. Nothing else, nobody else — see Privacy.