🌀
← Back to Articles & Artefacts
artefactswest

Synthesis: Six Fields, One Pattern

IAIP Research
alt2github-orchestrate-multiple-agent-communication-thru-git

Synthesis: Six Fields, One Pattern

The Convergence Pattern

Agent orchestration for workflow management sits at the intersection of six academic traditions:

  1. Multi-Agent Systems ← How agents coordinate
  2. Workflow Orchestration ← How tasks are sequenced (LangGraph model)
  3. Agent Communication Protocols ← How agents discover tools (MCP standard)
  4. Process Mining & BPM ← How workflows are logged and improved
  5. Agentic Planning ← How agents reason about dependencies
  6. Human-AI Teaming ← How humans and agents make decisions together

How MantisBT Fits

FieldMantisBT RoleMCP ExposureEngineering Implication
MASShared state ✓REST API → MCP toolsAgents read issue state as shared truth
Workflow Orch.Task graph storageIssue relationships (parent/child)Dependencies encoded in issue links
ACPService endpointNative REST + MCP serverAgents call via standard protocol
BPMExecution logAudit trail in notes/historyProcess mining possible if logs structured
PlanningDependency referenceQuery APIAgents walk parent→child chains
TeamingHuman-editable UI + agent APIDual interfaceHumans design workflows; agents execute them

Key Findings

  1. MantisBT's REST API + MCP server make it agent-readable without custom translation.
  2. Issue relationships (parent/child, depends-on) map cleanly to task dependency graphs used in LangGraph.
  3. GitHub integration via plugins links code commits to MantisBT issues, creating traceability from design → task → execution → code.
  4. Process mining becomes possible if execution logs are structured (currently partial; needs custom logging in notes or custom field).
  5. No bidirectional GitHub sync natively — workaround: custom field storing GitHub issue ID, triggered by webhook.

Engineering Implications

  • For agents: Query MantisBT REST API or MCP tools to get task state, dependencies, and progress. Safe assumption: parent issue holds decomposed subtasks; children are executable units.
  • For humans: Design workflows in MantisBT UI; agents execute them by reading and updating issue state.
  • For process improvement: Log execution traces (agent decisions, subtask completions) in issue notes or custom field; mine logs for workflow optimization.

Known Gaps

  1. Process mining tooling — MantisBT doesn't expose structured event logs; custom script needed to extract and analyze traces.
  2. Bidirectional GitHub sync — MantisBT → GitHub requires custom plugin or webhook; common workaround is a gh_issue_id custom field.
  3. Workflow sequence runner — MantisBT stores the state; you need LangGraph (or similar) to traverse and execute the graph.

Recommended Next Steps (from Perplexity Deep Search)

  1. "Supervisor agent pattern + MantisBT MCP" — Design how a manager agent uses MantisBT as task registry and LangGraph for execution control.
  2. "Process mining from MantisBT audit trails" — Extract structured logs and apply BPM discovery algorithms to auto-generate or validate workflow definitions.
  3. "Agent-native project management UI" — Evaluate whether MantisBT's UI + MCP layer is sufficient, or if a specialized agent-first tool (Plane, Linear, or custom) would be better long-term.