PlanWeaver

Writing, practicing, creating, trying to make the world a bit more interesting.

PlanWeaver is a collaborative Agent network for 'universal solution planning': users submit brief requirements (community events, business plans, project launches), and the network automatically discovers and teams appropriate skill Agents (outline, evidence & materials, process & budget, copywriting, review), producing in parallel before Reviewer triggers rework based on acceptance checklist, finally delivering deployable Markdown solution packages. Each collaboration accumulates reputation and capability profiles, making the network more accurate and efficient with use.

Architecture: Broker orchestration decomposition + Scout discovers available Agents + multiple Executors parallel production + Reviewer checklist-based acceptance and RETRY + SkillIndex records reputation and capability profiles. Protocol: Contract-style task assignment (Task Contract) and traceable materials (Evidence Cards), ensuring deliverables are verifiable, reusable, iterable. Output: Internal structured (JSON) → Final rendered as Markdown delivery package; end-to-end demo and interaction in OpenAgents Studio. Extensible: Connect retrieval/document/formatting tools as Mods, enhancing fact verification and format unification; supports different domain planning templates (business/academic/events).

Requirement parsing: Standardizes user Brief and extracts hard constraints/output specs (supports multi-scenario generalization). Smart discovery and teaming: Selects optimal Agent combinations based on skill profiles. Parallel division: Outline, evidence cards, process & budget, promotional copy produced in parallel. Review and rework loop: Reviewer checklist scoring, problem localization, RETRY instruction issuance. One-click delivery package: Generates proposal.md + appendix_evidence.md + changelog.md. Network learning: Records task performance and updates reputation/capability profiles, improving subsequent matching and quality. Studio demo script: 3-minute stable reproduction of 'discovery—collaboration—learning' loop.