Why we run agents interactively, not headless

Most agent tooling shells out to a one-shot, headless API call — the equivalent of claude -p "do the thing". It works in a demo, but it is fragile in production: no recovery when a step stalls, no shared context between phases, and a bill that scales with every retry.

The subscription you already pay for

trau drives a real, interactive session instead. It launches the same agent you use day to day, feeds it work, and steps back so the model keeps its context window warm across an entire task. No per-call metering, no surprise overage — just the plan you already have.

Headless is convenient until the first long task. Interactive is convenient forever.

What this unlocks

• -Long-running tasks survive transient stalls instead of failing the whole call.
• -Context carries across phases — plan, implement, verify — without re-priming.
• -Costs are predictable because you are not paying per token on every retry.

This post is a placeholder. The full write-up — including benchmarks against headless runs — lands here soon.