Agents
Declarations in configuration, not personalities in a prompt.
An agent here is a declaration: a name, the model it uses, the tools it is allowed to touch and how it is triggered. Templates live in version-controlled configuration, so you can read the agent surface of the system without running it.
The roster
The repository holds 55 agent templates. 43 agents were running in OpenFang on 27 October 2026; 37 of them are Claude Code subprocesses. Beyond the per-space agents, a set of general agents handles work that crosses domains.
| Agent | Handles |
|---|---|
| coder | Code generation and modification |
| debugger | Debugging |
| devops | Operations — currently also the agent Telegram messages go to |
| researcher | Web research and retrieval |
| planner | Decomposing a request into ordered steps |
| writer | Drafting and editing prose |
| knowledge | Knowledge retrieval |
| security | Security scans |
| orchestrator | Routing and coordination, without a model of its own |
| fallback | Anything that did not match |
Agents do not chat
Agents receive a mission, execute it, and return a result. When something needs several of them, the orchestration is explicit and recorded rather than emerging from a group conversation. This is deliberately less impressive in a demo and considerably easier to debug at three in the morning.
It also means the cost of a request is easier to bound. A system where agents talk until they agree has no upper bound on tokens. A system where an orchestrator dispatches a known set of missions does.
Capabilities are switches
Capabilities are enabled and disabled in the registry. When something is switched off, the file says so and the router skips it. Ten of the 119 capabilities are switched off today.