Internal / Hypothetical — Strategy Mock. This site is used for internal product positioning and fundraising preparation. Customers, logos, metrics, and financing described here are illustrative unless explicitly marked “verified.”
Agent swarm orchestration (internal)

Run a swarm. Keep control.

GasTown coordinates specialized agents, persistent work items, and merge-safe execution — so you can ship faster without losing accountability, budget, or clarity.

How we're different

  • Roles you can reason about (Mayor, Witness, Polecats, Refinery).
  • Work state lives outside the chat (Beads), so sessions stay disposable.
  • Merge queue + budget caps are first-class — not afterthoughts.
Trust rule: AI drafts changes; humans approve merges and high-risk actions.

What breaks at swarm scale

Design becomes the bottleneck

When code is cheap, clarity is expensive. GasTown keeps intent visible with queues, review gates, and decision logs.

Context rot

Long-running sessions degrade. GasTown externalizes state into Beads and treats sessions as disposable workers.

Merge conflicts + drift

Parallel work creates collisions. GasTown uses a Refinery merge queue and staged diffs to keep changes small and landable.

The town roles

Mayor concierge

Human-facing coordinator. Decomposes goals into Beads, assigns roles, and keeps the plan coherent.

Witness supervisor

Unblocks workers, enforces checklists, and requests clarifications when intent is underspecified.

Polecats workers

Disposable specialist sessions that do one job (add tests, fix a bug, refactor a module) and report back.

Refinery merge queue

Stages diffs, resolves conflicts, and ensures every change is verified before landing.

Archivist audit

Maintains an immutable trace: who did what, what evidence exists, and what was approved.

Boot (the dog) nudges

Keeps the swarm fed. Heartbeats workers to check their queues and escalates stalled Beads.

Beads + seancing

Beads = persistent units of work

Each bead is a small JSON task with an owner, status, evidence links, and acceptance criteria.

Principle: keep durable truth outside the chat.

Seancing = restart without losing state

Spin up a fresh session and have it interrogate the last run: what happened, what’s stuck, what’s next.

Principle: sessions are disposable; Beads are durable.

Queues + heartbeat

Hooks keep agents moving

Every role has a queue and a “next hook”. Workers pull the next bead and publish progress continuously.

Guardrails prevent runaway loops

Budgets, retry limits, and review gates stop expensive thrash and make work predictable.

Example workflow: fix a production bug

  1. Mayor creates Beads: reproduce, fix, add tests, document.
  2. Polecats implement small diffs; Witness reviews against acceptance criteria.
  3. Refinery stages diffs, runs the verification harness, and lands changes in order.
  4. Archivist records evidence: logs, screenshots, test runs, approvals.

Start a pilot (internal)

We’ll configure roles, hooks, budgets, and a verification harness for one repo and one workflow.