galleon labs

We build agents that hold real jobs.

Three classes of agent, all in production. Each owns work that used to need a person, and has to clear a stated bar before calling the job done.

OneshotSWE agents
Davy JonesSocial agents
BarbossaOperations agents
3In production
SWE agents

Oneshot

Private

Product engineering is mostly an unglamorous loop: triage, reproduce, fix, test, review, ship.

Runs that loop inside Slack. A request, a GitHub event, or a scheduled check goes in; a verified deploy comes out, reported in the thread it started in.

  1. Slack-native

    Thread-scoped context, streamed progress, cancellable work.

  2. Quiet by default

    Event-driven checks on PRs, drift, and incidents. Most end silently.

  3. Skill-routed

    Every job loads its own written procedure, not one long prompt.

  4. Gated on evidence

    Merge, deploy, and working software are three separate gates.

Building Davy Jones, Our AI InternThe write-uponeshot-shipThe CLI, on npm
Social agents

Davy Jones

Private

Agentic mirroring: with enough context, can an agent become indistinguishable from its owner?

Fed years of Andrew's writing, decisions, and obligations, then given somewhere to use them. Outward actions and money stay gated.

  1. A real corpus

    Years of writing, decisions, and obligations. Not a persona prompt.

  2. His own identity

    Email, phone number, social accounts, browser, and a funded wallet.

  3. Acts unprompted

    Reads his own channels, keeps notes, and starts work unasked.

  4. Outlives the model

    The identity and its evidence survive every model swap.

The threadHow it works@davyjonesinternPosting for himself
Operations agents

Barbossa

Private

Running infrastructure and a fleet of agents stops being possible by hand around the third system.

Runs the servers and supervises the agent teams on top of them. Three teams today, with a stated health contract for every asset.

  1. One register

    Teams, lanes, schedules, and state declared once, read from one surface.

  2. Product-owned truth

    Each repository keeps its own code, deploys, and decisions.

  3. Health to outcome

    Checks, investigations, and releases form one chain, not scattered crons.

  4. Bounded authority

    Personal and work systems stay apart. Human calls stay human.

barbossa-devEarlier generation · MITbarbossa.devDocs for that versionSetting up BarbossaWrite-up, first generation

The products are the evidence.

Peerlytics and USDCtoFiat run on the ZKP2P market. KnowFulham maps west London. All of them are built and maintained with the systems above.

See the work