TALOS

Workflows

How we build with coding agents.

Each of us works in one agent conversation. The agent on the laptop plans and decides, cheaper models take the work that can be checked, outposts run the heavy jobs on remote GPUs, and a session can be handed to a teammate who carries on where it stopped. Below is the loop, then one page for each tool we built for it.

The loop

One conversation is the brain. Everything else is a worker or an outpost.

The main agent runs on Claude, on the laptop of whoever is driving. It reads the code, writes the plan and reviews every result. Cheaper models only take tasks with a clear check, and remote machines only execute. Choose a step to see where it happens.

The loop, step by step. One conversation on the laptop; each step reaches out from it.
Your laptop
Claude Codemain agent on Claude · plans, decides, reviews
plan.mdapprove · review
Youapprove the plan · review what comes back
claude_xlauncher + local gateway
every model callrouted by model
Model providersAnthropic for the main agent · a cheaper model for workers
talos-skytsky CLI + MCP server
sshjobs · events · previews
Your machinesrented GPUs · cloud VMs · lab servers, each with an outpost
easy_harborin design
bundleone file
A teammatereads, replays or continues the session in their own Claude Code
  1. The main agent reads the code and writes one plan file, then keeps it current as work lands. A person approves the plan before anything runs, and reviews every result that comes back.

  2. Mechanical work with a clear check goes to workers on a cheaper model, through claude_x. They run as ordinary Claude Code subagents, and the main agent reviews what they return. How claude_x works

  3. Training, rendering and simulation run on GPU machines through talos-sky outposts. Job ends, log alerts, renders and metrics come back into the same conversation. How talos-sky works

  4. When a teammate should carry on, the session becomes one bundle they can read, replay or continue in their own Claude Code. The easy_harbor design

The tools

One tool for each step that leaves the laptop. Each has its own page: what it does, a step-by-step diagram of how it works, and where it is going.

How we work

Plan first
Every task starts as one plan file that the agent keeps current. A person approves it before work starts and reads it again before anything ships.
Parallel, then review
Independent changes run as subagents in separate git worktrees. The main agent reviews each one and merges them in a fixed order. talos-sky 0.4 shipped this way.
Cheap work, checked
Workers get well-specified tasks with a clear done check, such as tests that must pass or citations to spot-check. Claude reviews what comes back before anyone relies on it.
Rent small, return fast
The cheapest GPU that fits the job, rented only after a person confirms the offer, and destroyed when the job is done.
Ask before it leaves
Pushes, deploys, rentals and anything sent outside the machine are confirmed by a person, every time.
Tools for people too
Every command has help and examples, every error says what to do next, and every agent action has a terminal equivalent.