Nautilus is the runtime Truline uses to run coding agents against real repositories. Each worker gets a capsule: the task, the skills it must load in order, the files it may write, and the evidence it must return. The runtime admits the capsule before the model starts, holds the write scope while it runs, and records every run.
› nautilus run --task ship-fix --model glmadmission capsule truline-ship-fix · release c8334f7ainputs task.md 9f2c… context.md 41ab… (2 hashed)charts precise → tdd → code-review (all read this session)scope write: src/lib/parse.ts, tests/parse.test.tsadmitted run 3b1e-… startedworker reading tests/parse.test.tsworker failing test written: parses trailing commaworker edit src/lib/parse.ts (+11 −2)check npm test ✓ 84 passedcheck lint ✓review route: astra · findings: 0 blockingreceipt outputs 2 · refusals 0 · 4m12s› _
Nothing bypasses admission: no raw model binary, no shell escape, no nested worker. A capsule that declares a skill the session never read is refused before any file is touched. A worker that can't do its job inside its scope reports the exact refusal and leaves the job open. The parent never quietly finishes it with its own tools.
A calibrated judge scores each closed decision, like which skill applies to a task or whether a heading states its conclusion. Every question and threshold lives in one file, next to the eval that set it. A pass won by weakening a check certifies nothing, so the runtime refuses it.
Each run writes its capsule, its inputs by hash, the tools it used, the refusals it met and the outputs it made. Rendered work is reviewed by a different model route from the one that made it. The record is what any claim of "done" is measured against.
The engineering method behind Nautilus is written up in the Agent Engineering Handbook: the guides, the ideas and the failure cases behind each rule.
Read the Agent Engineering HandbookThe runtime Truline uses to run coding agents on real repositories. Each worker runs inside an admitted capsule with a set task, skills, write scope and required evidence.
It reports the exact refusal and leaves the task open. The parent does not quietly finish it with its own tools.
Every run writes a record: its capsule, its inputs by hash, the tools it used, its refusals and its outputs. Claims of completion are checked against that record.
In the Agent Engineering Handbook, which covers the guides, ideas and failure cases behind each rule.
Bring one agent workflow you want to run safely. We'll show how a capsule, a write scope and a run record would apply to it. Want the short version first? Read our pages on Claude Code rules and AI agent governance.
A slow week shows where the time goes, then the next busy week buries it. A better process gets written down, then nobody uses it. The lesson was learned. It never became how the business runs.
Every part of a business can run as a closed loop: admin, sales, marketing, client service, hiring and training. Market Loop is software plus AI that connects to your business and the systems it already uses, and closes the gap between knowledge and implementation. Each lesson becomes a change you approve. The change goes in, the result gets measured, and that starts the next round.