Agents that can carry the watch

The useful agent is already inside the work

Most agent demonstrations begin with a blank chat box. Real work rarely does. It begins in a Slack thread, a pull request, a failing production journey, a product queue, or a responsibility that has quietly gone stale.

Galleon's agents are built around that difference. They enter through the surfaces where the work already exists, inherit a bounded operating context, select a specialised procedure, and return with evidence that something changed. The conversation is the cockpit, not the product.

Davy Jones is the experiment in becoming Andrew

Davy Jones asks how close an agent can get to becoming its owner as the underlying models improve. It is supplied with Andrew's writing, decisions, working memory, obligations, preferences, personal tools, and the boundaries that preserve his authorship. It also has continuity outside the prompt: email, a phone number, social and publishing accounts, a browser desk, and a funded wallet for bounded machine payments.

The point is not to make a chatbot that sounds vaguely like Andrew. It is to keep testing whether a frontier model can hold more of a real person's context, judgement, initiative, and practical interfaces over time—without confusing simulation with permission. The identity stays; the harness and model can change.

Oneshot removes the repetitive layer at work

Oneshot is Andrew's Slack-native work agent. He built it to automate the repetitive operational layer around software work and help the wider team ship. It responds in DMs and mentions, maintains thread-scoped sessions, streams compact progress, exposes a private App Home, and can turn a handover in conversation into a managed piece of repository work.

The stronger part is what happens without a fresh prompt. Bounded watches respond to eligible GitHub events, critical product journeys, repository drift, unanswered team responsibilities, and the health of adjacent systems. A large curated skill catalogue gives each job its own procedure instead of asking one swollen system prompt to improvise everything.

This is what makes the agent feel like a reflection of its owner. It is not trained to imitate Andrew's sentences. It is designed to share his operating instincts: find the real source of truth, distinguish a merge from a deployment, keep authority explicit, finish the loop, and stay quiet when there is no result worth interrupting someone for.

• Slack is the native control plane, not a notification afterthought.
• Proactive loops are narrow, observable, and tied to real responsibilities.
• Skills carry detailed procedures; the core prompt carries principles and boundaries.
• Tests, review, deployment identity, and live behaviour are separate proof gates.

Barbossa keeps the personal fleet legible

Barbossa is Galleon's personal operations runtime. It currently supervises KnowShip, Shipwright, and USDCtoFiat Growth, and is being extended to cover the health of the wider product fleet. The products keep their own repositories and domain truth; Barbossa owns shared scheduling, state, investigations, release lanes, health, and the fleet console.

That distinction prevents the usual agent-fleet failure: a folder of unrelated cron jobs that are individually alive but collectively impossible to reason about. Barbossa declares each team and lane once, records what happened, exposes what needs attention, and connects detection to investigation, work, release, and outcome review.

One philosophy, distinct jobs and trust boundaries

Davy, Oneshot, and Barbossa share a standard but not a single identity, job, or pool of secrets. Davy is the personal AI-self experiment. Oneshot works inside Andrew's professional Slack and repositories. Barbossa operates the personal Galleon product fleet. The separation is a feature: a capable agent should know exactly which ship it is aboard.

The aim is not maximum autonomy in every direction. It is useful autonomy inside a legible contract. A strong agent knows what it owns, what evidence closes the loop, what remains a human decision, and when silence is the correct result.