A multi-agent orchestration experiment: role-based agents composed into pipelines.
Built early in the agent-tooling cycle. Researcher → Synthesizer → Critic → Writer, each with its own system prompt, all coordinated by a small orchestration layer. A working artefact more than a product.
In one paragraph.
Claude Colony is a small framework for composing role-based agents into pipelines. Each "colony member" is a Claude call with its own system prompt, memory window, and output schema. The orchestrator routes work between them and handles retries, fallbacks, and human approval gates.
Built before agent frameworks were mainstream. The mental model — agents as named roles with strict contracts between them — held up well; it’s the same model the Drops agent uses today.
The motivation.
I wanted to test whether composing small, focused agents would outperform one large agent doing everything. Spoiler: yes, when the roles are narrow enough and the contracts between them are strict.
The features that matter.
Role-based agent definitions
Each role is a system prompt + output schema. Strict contracts between members.
Pipeline orchestrator
Routes work, handles retries, surfaces approval gates.
Sample colonies
A research colony, a writing colony, an evaluation colony. Forkable patterns.