Sprint 7 closed: Engineering Agents

Sprint 7 is closed.

The goal was to introduce specialized engineering agents — architecture, code review, documentation, debugging, and onboarding — built on skills, memory, scanner output, and MCP. Agents do not call external LLMs; they gather project evidence and return structured briefings so the host model can execute the work with the right context and checklists.


What Sprint 7 delivered

Every item on the roadmap is checked off:

  • Architecture agent — ADRs, repo layout, dependency boundaries
  • Code review agent — git diff stat, optional GitHub PR context
  • Documentation agent — doc gaps, GitHub doc draft excerpts
  • Debugging agent — symptom-driven investigation with recent commits and diff
  • Onboarding agent — vision, sprint, ADRs, and first safe commands

Product MCP is now v1.2.0 with agent_list and agent_run. Local MCP is v1.18.0 with the same agent surface.


New package: @meronq/agents

Module Role
listAgents() / getAgent() Built-in agents (v0 — no project overrides yet)
runAgent() Gather context and return AgentBriefing
gatherAgentContext() Domain-specific context (architecture, PR, docs, debug, onboarding)

Each briefing includes:

  • context — scanner index, git, GitHub, doc excerpts
  • checklist — ordered review or investigation steps
  • systemGuidance — role instructions for the host LLM
  • recommendedSkills — e.g. project.test, adr.create
  • suggestedPrompts — example user intents

Execution context lives in @meronq/server-core (createAgentContextAsync) — same pattern as skills, without circular dependencies.


Surfaces

meronq agent list
meronq agent run onboarding
meronq agent run debugging --symptom "pnpm test fails in @meronq/skills"
meronq agent run code_review --pr 12 --focus security

MCP tools: agent_list, agent_run. Handshake returns available_agents[] alongside available_skills[].

agent_run is read-only in v0 — no runtime or file-write permission gates. Agents recommend skills; skills execute procedures.


Pipeline

Intent (agent name + inputs)
  → getAgent
  → createAgentContextAsync (index, git, GitHub, docs)
  → gatherAgentContext (domain-specific)
  → AgentBriefing (context + checklist + guidance)
  → Host LLM + recommended skills

Skills answer how to run build, test, or create an ADR. Agents answer when and why, with the evidence an assistant needs before acting.


Sprint 8 — Productization

Sprint 8 goal: turn prototypes into a usable product — desktop app exploration, cloud backend, team workspace, authentication, and billing.

Agents and skills are the intelligence layer. Productization makes that layer accessible beyond local MCP and CLI.


Follow along

Sprint 6 made procedures reusable. Sprint 7 gives AI clients specialized engineering roles with evidence-backed briefings.

← Back to blog