Software Engineering Collective

Tools / 01 · SI · Multi-agent orchestration

Agent Swarm

Status: Open source · Active development

Agent Swarm is an open-source multi-agent system that coordinates 15 specialised SI agents across the full software development lifecycle, from planning and build through verification, release and maintenance. An orchestrator delegates to specialists, every artifact has a single writer, quality gates fail closed and stalled work escalates to a human. Built by Spectrum Web Co in Sydney.

Agent Swarm turns a business request into a planned, built, reviewed and released change, with every step traceable back to the request. Each agent owns one kind of artifact, each handoff is signed and checked, autonomy is capped by runtime policy, and a human steps in the moment work stops converging.

View the repo on GitHubTalk to us about agentic delivery

About

What is Agent Swarm?

Agent Swarm is the SDLC swarm inside KanbanOS, released as open source. It covers the whole lifecycle in one governed system, from requirements and design through backend, frontend and data engineering to QA, code review, security, DevOps, release, observability, maintenance and documentation, with correlation tracking from business request to release.

  • 15 specialised agents under the A01 Orchestrator
  • Single-writer artifacts and fail-closed gates
  • Autonomy levels L0–L4 enforced at runtime

01 — The team

One orchestrator, fourteen specialists.

The A01 Orchestrator plans each request and delegates to specialist agents for requirements, design, engineering, QA, code review, security, DevOps, release management, observability, maintenance and documentation. Each specialist owns its own artifacts, independent work runs in parallel, and correlation tracking ties every output back to the business request that started it.

  • Single-writer artifact ownership
  • Parallel execution and autonomous delegation
  • Correlation tracking from request to release

02 — The gates

Uncertain work stops.

Every handoff is a signed swarm.v1 message and every quality gate fails closed: if an agent cannot prove a result, the work does not advance. Failed work gets at most two automatic retries, then escalates to a human. Autonomy is set per run from L0 to L4 and enforced by runtime policy, not by prompt alone.

  • At most two automatic retries, then a human
  • Signed swarm.v1 protocol messages
  • Humans approve merges and releases

03 — Run it

Simulate first, then execute.

Orchestration runs from the command line with plan and run scripts, through Claude Code subagents or via Trae SOLO agent registration. Dry-run simulation shows the plan before anything executes, headless execution handles real runs, and task state lives in an inspectable SQLite state machine. The repository ships architecture docs, a seven-level validation protocol and deployment and scalability guides.

  • Dry-run simulation and headless execution
  • Claude Code subagents with Python tools
  • Seven-level validation test protocol

04 — Traceability

From request to release, one thread.

Correlation tracking follows work from the business request that started it through to release. Because each artifact has a single writer and each handoff is a signed message, you can see which agent produced what and trace any release back to the request behind it, without reconstructing the story from terminal logs.

  • Correlation from business request to release
  • Signed, attributable handoffs
  • One owner per artifact

Who it’s for

Built for engineering teams that ship with agents.

Agent Swarm is for engineering organisations that want parallel SI agents across the software lifecycle without giving up code review, traceability or human control over what ships.

  • Teams already running SI coding agents
  • Platform and developer-experience teams
  • Engineering leaders who need auditable agent governance
  • Studios shipping several products at once

FAQ

Agent Swarm, answered.

01

How does Agent Swarm work?

Agent Swarm works as a hierarchy of 15 specialised SI agents that take a business request through the full software development lifecycle: plan, build, verify, ship and maintain. The A01 Orchestrator plans the work and delegates to specialists covering requirements, design, backend, frontend and data engineering, QA, code review, security, DevOps, release management, observability, maintenance and documentation. Independent tasks run in parallel, delegation happens autonomously within policy, and correlation tracking links every artifact back to the original business request, so any release can be traced to the reason it was built.

02

How does Agent Swarm stop unverified SI code from shipping?

Agent Swarm stops unverified code with fail-closed quality gates. Every artifact has a single writer, so each one has exactly one accountable agent and is never overwritten by another. Every handoff is a signed message, and a gate opens only when the result can be proven; if it cannot, the work does not advance. Failed work enters a bounded rework loop of at most two automatic retries, after which it escalates to a human instead of looping or guessing. Dedicated code review and security agents are part of the swarm itself, not a separate step bolted on afterwards.

03

How autonomous is Agent Swarm, and where do humans approve?

Agent Swarm’s autonomy is set per run on five levels, L0 to L4, from fully supervised to largely autonomous, and the level is enforced by runtime policy rather than by prompt instructions an agent could ignore. Teams can start supervised, observe how the swarm behaves on their codebase and raise autonomy as trust is earned. Humans keep merge and release approval, and any work still failing after two automatic retries is escalated to a person. That gives engineering leaders a governance model they can explain to security, risk and audit stakeholders.

04

How does Agent Swarm integrate with Claude Code?

Agent Swarm runs its agents as Claude Code subagents, each defined with XML-tagged prompts and backed by Python tools. Agents communicate through a signed message protocol, swarm.v1, and task state is held in a SQLite-backed state machine you can inspect at any point. Orchestration is driven from the command line with plan and run scripts, supporting dry-run simulation to preview a plan and headless execution for unattended runs. Agents can also be registered with Trae SOLO. The swarm runs on Agent Substrate and is the SDLC swarm inside KanbanOS.

05

Is Agent Swarm open source and ready for production teams?

Agent Swarm is open source on GitHub and in active development. The repository ships more than code: architecture documentation, a seven-level validation test protocol, and deployment and scalability guides, so a platform team can evaluate how it behaves before trusting it with real work. A sensible adoption path is to start with dry-run simulation, run at a low autonomy level on a contained repository, and raise autonomy only as the gates prove themselves on your codebase. Teams that want it wired into their own delivery workflow can talk to Spectrum Web Co about Ship a Product.