luigidemasi opened a new issue, #33: URL: https://github.com/apache/camel-kit/issues/33
> Imported from [luigidemasi/camel-kit#60](https://github.com/luigidemasi/camel-kit/issues/60), opened by [luigidemasi](https://github.com/luigidemasi) on 2026-05-10. > Closed by [luigidemasi/camel-kit#62](https://github.com/luigidemasi/camel-kit/pull/62), merged on 2026-05-12. --- ## Description Expand subagent usage across camel-kit skills to isolate review traces, MCP responses, validation output, and knowledge queries from the main conversation context. Currently, spec-reviewer and quality-reviewer run as guides within the main context, and all MCP/verification output accumulates in the orchestrator's conversation. The goal is to make the main context a thin orchestrator that dispatches domain-specific work to subagents and receives only structured summaries back. Based on analysis in the [Agent Skills Context Engineering](https://github.com/addyosmani/agent-skills) deep dive. ## Areas to Expand 1. **Formalize reviewers as standalone agents** — extract `spec-reviewer-criteria.md` and `quality-reviewer-criteria.md` into `agents/` as reusable persona definitions. Each runs in its own context and returns only findings. 2. **Fan-out in `camel-ship`** — run spec, quality, and security reviewers in parallel rather than sequentially. Main agent merges reports. Keeps review traces out of main context. 3. **Validation as subagent** — when `camel-execute` triggers `camel-verify`, run verification in a subagent and return only the structured report. 4. **MCP lookups as subagent** — catalog verification calls (Iron Law 1) batched into a research-isolation subagent returning a summary (e.g., "all 5 components verified, 1 unknown: `camel-xyz`"). 5. **Knowledge queries as subagent** — `camel-knowledge` queries run in an Explore-type subagent, returning the relevant answer without loading full search results into main context. ## Three-Layer Composition Model | Layer | What it is | Today | Target | |-------|-----------|-------|--------| | Skill | Workflow with steps and exit criteria | SKILL.md files | Keep as-is | | Persona | Role with perspective and output format | Inline in skills | Extract to `agents/` as standalone personas | | Command | User-facing entry point | `/camel-*` commands | Keep as-is | ## Design Constraints - **Personas do not invoke other personas.** Composition is the job of slash commands or the user. - Each persona must have a `## Composition` block stating: invoke directly when / invoked via / do not invoke from another persona. - Max depth = 1: command → persona (no deep persona trees). - Follows endorsed patterns: parallel fan-out with merge (reviewers), research isolation (MCP/knowledge), direct invocation (validation). ## Implementation Steps - [ ] Extract spec-reviewer and quality-reviewer criteria into standalone persona definitions in `agents/` - [ ] Add `## Composition` blocks to each persona following agent-skills' pattern - [ ] Prototype subagent fan-out in `camel-ship` — run reviewers in parallel and measure context savings - [ ] Move MCP catalog verification to research-isolation subagent - [ ] Implement in-flight doubt cycle in `camel-execute`: CLAIM → EXTRACT → DOUBT mini-cycle for non-trivial subagent decisions, with 3-cycle hard cap ## Context Savings Each subagent has its own context window. Only structured summaries flow back to the orchestrator. For a typical 5-task plan with 2-stage review per task, this could prevent ~60-70% of review-related tokens from accumulating in the main conversation. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
