→ blog

Field notes

Release stories, agent workflow deep dives and measurements from running an autonomous delivery loop on real tickets.

Entries
007
First filed
Last filed
RSS feed
LOG-007AI Insights
9 min read

We stopped asking LLMs for small decisions. A case study with TypeSafe's Jev.

How trau uses Jev, TypeSafe's System One model, for the small judgments inside an autonomous coding pipeline: ticket complexity, model routing, lesson recall, duplicate hints and stall detection. What worked, what it costs, and the rough edges.

Read entry

Earlier entries

  1. LOG-006AI Insights7 min

    We audited our agent pipeline against the new rules of context engineering. One rule we broke on purpose.

    A week spent measuring every context surface trau assembles. Prompt text was never the cost, permission-style wording gets 0% uptake unattended, and auto-memory quietly broke cold verification.

  2. LOG-005News2 min

    Trau 2.0: the hub

    A CLI grew a web app, and then the web app ate the product. Why the loop needed a persistent surface — and what stays terminal-first.

  3. LOG-004AI Insights1 min

    Why we run agents interactively, not headless

    Metered -p API calls are fragile and expensive. Here is the case for driving real interactive sessions on the plan you already pay for.

  4. LOG-003Engineering1 min

    Loop mode: how trau fans a ticket out to its children

    A look at the detection mechanism that decides when a task runs solo and when it spawns a labeled, ready-for-agent loop.

  5. LOG-002AI Insights1 min

    Counting work like a human, not a stopwatch

    Wall-clock time lies about effort. We explain how trau correlates machine seconds to human-equivalent work for standups and invoices.

  6. LOG-001Guide1 min

    Picking the right model for each phase of a task

    Claude, Codex, or Kimi — and at what effort? A practical guide to per-phase routing in .trau.ini for cost and quality.

(END)You’ve reached LOG-001, where the log starts.