Kyoyo turns the agent setups scattered across your engineers' laptops into one governed line: your workflows, your gates, your context, with every feature costed and reviewed before it ships.
Agents made individual engineers fast. They also gave every engineer a private setup: their own prompts, rules, MCP servers and habits. The organization got faster and less consistent at the same time.
Two engineers, two setups, two different standards for tests, review and rollout.
What works lives in one laptop's config. It doesn't spread, and nobody reviews it.
Finance sees vendor invoices. Engineering sees PRs. Nothing joins the two.
A factory is not the absence of skill. It's where the best way to do a job is written down, improved by the people who do it, and used by everyone.
| Workshop · each laptop | Factory · Kyoyo | |
|---|---|---|
| Setup | A private set of prompts, rules and tools per engineer | Versioned workflows per change type, shared by every team |
| Quality bar | Whatever the author's setup enforces | Rules and gates on every run, the same for everyone |
| Review | Whatever the author remembers to ask for | A human gate before every merge, deploy and migration |
| Context | Gone at the end of the session | Recorded on every run, queryable by API and MCP |
| Cost | One invoice per vendor, no unit of work | Cost per feature, per stage, per workflow |
| Improvement | One person tunes their own config | Anyone proposes a workflow change; approved changes ship to all |
| Audit | Reconstructed from Slack and git log | Every agent action and approval, exported to your SIEM |
Every workflow mixes deterministic steps you already trust with agent steps where judgment pays off. Nothing irreversible happens until a person signs.
09:14:02 · Trigger LIN-2291 labelled bug · run R-4471 started 09:14:05 · Rule /VALIDATE passed · repro steps, owner, severity present ✓ 09:16:40 · Agent reproduce wrote failing spec invoice_timezone_spec.rb ✓ 09:21:12 · Agent draft-fix opened PR #1291 · 3 files · +42 −9 ✓ 09:27:30 · Rule /CI ran 412 checks · 0 failed ✓ 09:27:31 · Decision risk high? → false · no schema or production-data change 09:27:31 · Gate owner review → waiting on billing-owners ● · Next: call workflow release v6 · canary, watch, promote · Run cost so far $3.84 · 13 min elapsed
Moved invoice dates to UTC so exports match the ledger. Added a failing spec first; it passes now. Needs your review.
Rules, decisions, agent steps and human gates on one canvas. Your CI and approvals become nodes, not afterthoughts.
Every run records the chain from ticket to spec to PR to deploy to outcome. The next run starts from it.
Runs happen in the cloud, not on laptops, with the controls an enterprise expects. Every run exports to systems you already trust.
Tokens, compute and review minutes rolled up from run to PR to feature, with quality numbers beside them.
Spend per engineer doesn't tell you what you bought. Kyoyo measures the unit that matters: a shipped feature, and what it took to get there.
Every line reports into one view. Six questions a VP of Engineering asks every week, each answered by its own dashboard.
Which lines are running, and which are stuck?
Where does work wait, and for how long?
Which teams are holding the line, and how long has it waited?
What got past the gates, and which step let it through?
What does each line cost per feature, against budget?
Which shared workflows carry the most lines?
| Line | In flight | At gates | Shipped | FPY | $/feature |
|---|---|---|---|---|---|
| Bug fix v14 | 42 | ● 9 | 168 | 78% | $96 |
| Feature v9 | 61 | ● 7 | 94 | 64% | $412 |
| Dependency upgrade v7 | 51 | ● 1 | 121 | 88% | $38 |
| Incident follow-up v2 | 12 | ● 2 | 18 | 69% | $205 |
| Schema migration v3 | 8 | ● 4 | 11 | 52% | $730 |
| ▣ Release v6 called by 4 lines | 12 | 0 | 412 | 97% | $9 |
Security's oldest item has waited a day: a schema migration touching patient billing data.
Lead time 6h 14m, of which 2h 5m is work. The review gate's wait is 59% of it.
Two causes account for two thirds of escapes. Both are fixed in the workflow, once, for every team.
Policy, CI and approvals are rules, not prompts. Agents work between them.
Merge, deploy, migrate and send all pass a human gate. Every time.
Agents run like processes: a status, a log and a stop button, attributable and reversible until someone commits.
Cost and quality per shipped change, not seats and suggestions accepted.
許容 (kyoyō) is Japanese for tolerance: what is permitted, within limits. On a shop floor it names the allowed variation in a part, 許容差, the gap between the drawing and what can still ship.
That is the idea behind the product. Agents can take their own path to a change. The change still has to land inside the limits your team sets: your rules, your gates, your review. Inside tolerance, it ships. Outside it, the line stops and a person decides.
The name also nods to the Toyota Production System, which shaped how this factory thinks about quality: stop the line on a defect, and let a person decide.
Pricing is per factory, a platform fee plus usage, not per seat. We scope it with you.