Heap (specialists)
When the chat uses heap mode, Agentron runs at Level 4 (recursive/self-building): a planner chooses an ordered set of specialists and can create new tools and agents on the fly. Each specialist has a limited set of tools (up to 10); the system runs them in order (or in parallel where allowed) and merges context. This keeps responses focused and within token limits.
Production heap (interactive)
The diagram below shows the current production heap: top-level specialists and their delegates. You can pan, zoom, and drag nodes. This snapshot is generated at docs build time from the same registry the app uses.
Loading heap…
Top-level specialists
| Specialist | Purpose |
|---|---|
| general | Conversation, ask user, format response, remember, list agents/tools |
| workflow | Workflows: list, get, create, update, execute, runs, cancel, respond, versions, rollback |
| agent | Agents: CRUD, versions, rollback, LLM providers; OpenClaw (send, history, abort) |
| tools | Create and improve tools (HTTP, MCP, code); custom functions |
| improve_run | Improve current run/session only (no DB); observe run/feedback, session overrides |
| improve_heap | Improve the heap (registry, planner); register/update specialists |
| improve_agents_workflows | Improve workflow agents and workflows (studio DB); observe runs/feedback, update agents/workflows, training pipeline |
Some specialists delegate to sub-specialists (e.g. agent → agent_lifecycle, agent_openclaw) so that each node still has at most 10 tools.
See also
- Agentron (Chat): Where to find the chat and how it uses tools
- How Chat routes (heap): Planner and specialists in one sentence
Last updated on