Progressive planning in practice

Observed in a real Fleet

These four separate outcomes were observed in one authorized Fleet on 31 July 2026. They are faithfully shortened, not composited. Identifying details were replaced while sequence, ownership, authority boundaries and measured results were preserved.

Turn an ambitious strategy into honest evidence

One Captain outcome asked the Fleet to chart an evidence-backed long-range company strategy. The resulting graph held 15 durable Tasks: 2 completed, 2 active, 3 blocked and 8 waiting on dependencies. First Mate chose the smallest honest initial split: one Scout for the private company baseline and one for the public market landscape.

The public Scout returned a 43,638-character report grounded in 57 public sources. It ranked candidate openings, attached confidence to each and named evidence that would falsify them. It also refused to make the final company-specific choice because customer, pipeline, product-usage, cost, channel and capability evidence remained outside its authority. The Fleet made those missing inputs durable instead of hiding the gap behind a confident recommendation.

Repair, prove and roll out a production path

Two linked Tasks kept implementation proof separate from production authority:

1. AI Gateway repair Task. The owner added bounded, privacy-safe stream-failure telemetry and prevented cancellation or cleanup from masking the provider outcome. An authenticated long-stream test reproduced 384 SSE events and 798,994 bytes byte for byte.
2. Separately authorized release/rollout Task. A separate owner, with its own Captain authorization, published an immutable release and replaced only the Gateway. Durable storage, routing state and active Agent identities remained intact. A corrected production request completed with 2,480,890 bytes across 7,003 chunks and 8,721 SSE events. One persistent Mate completed a canary in its existing native session; broader rollout remained separately gated.

Upgrade persistent workers without erasing the company

Another outcome upgraded five persistent Mates serially. Each replacement retained its Agent identity, storage, native session, active Tasks and Assignments, private memory and child-workload boundary. The operator replaced one Mate at a time and would stop at the first unverified boundary. All five reached the reviewed release and reported evidence upward. Credentials, charters, ownership and unrelated external systems stayed untouched.

Avoid work that evidence made unnecessary

One accepted outcome asked for an approved CLI in the default Mate image. Inspection proved that an equivalent reviewed change was already in the published release, so coordination closed with no implementation and no competing pull request. Live rollout remained a separate Captain-gated outcome.

Together, the outcomes show the company loop: decide where to go, repair and ship the product, operate persistent workers safely and avoid unnecessary work.