Intent Understanding: Why This Packet Exists

Structural Tension

Current Reality:
Agent teams orchestrate work via ad-hoc dispatch patterns; workflow state is fragmented across logs, tickets, and memory. Human designers discuss workflow intent in prose; agents read unstructured data or make assumptions about task dependencies.

Desired Outcome:
Agents operate on a shared, queryable workflow state store (MantisBT + MCP) where task graphs, dependencies, and execution history are legible to both humans and agents. Design conversations and agent operations are grounded in the same academic frameworks.

Why Now

The convergence of three fields — workflow orchestration (LangGraph), agent communication (MCP standardization), and process mining (BPM → LLM fusion) — creates an opportunity to build agent-native workflow management on recognized academic foundations rather than reinventing patterns.

Intended Audiences

1. Architects — understand which academic fields ground each design choice
2. Agent Developers — know what state to query from MantisBT and how it maps to workflow concepts
3. Cross-Repo Stewards — reference this packet when evaluating agent orchestration decisions in Miadi, coaia-agent, or IAIP

Decisions This Packet Strengthens

• Should MantisBT be the state store? — Yes; it aligns with BPM and process mining traditions
• Which protocol for agent-MantisBT access? — MCP; it's becoming the standard for agent-to-service communication
• How to encode workflow dependencies? — MantisBT relationships (parent/child, related); readable by agents via REST API
• What success looks like — Agents traverse dependency graphs, query task state, and update issue notes without domain-specific glue code

Future Synthesis Points

• RISE framework specs for MantisBT integration (→ rispecs/)
• Process mining queries on MantisBT execution logs (gap: needs custom plugin or script)
• Supervisor agent that manages sub-agents via MantisBT task assignment (→ CLAUDES experiment)