Decisions

  • How autonomous should the Tech Lead be? (pick from planned vs. wait for assignment) · Phase 1 · ready
  • Git branching strategy (per feature vs. per task) · Phase 1 · ready
  • Escalation threshold (when agents ask you vs. decide themselves) · Phase 1 · ready
  • Report verbosity (always full vs. only non-trivial) · Phase 2 · ready

User Tasks

  • Review and approve design: agency-company-structure · Phase 1 · ready
  • Decide on model routing per role (Opus/Sonnet/Haiku) · Phase 1 · ready

Summary

Model Opus as a company: agents have roles, a reporting hierarchy, and produce structured reports. You are management.

Problem / Motivation

The autonomous coding agency (Group 9) defines what agents do technically, but not how they communicate. Without structured reporting:

  • You have no visibility into what agents did or why
  • Agents can’t coordinate or hand off work
  • There’s no audit trail beyond git history
  • You’d have to read diffs to understand what happened

The company metaphor solves this: defined roles, chain of command, structured reports.

Proposed Solution

Three agent roles (Tech Lead, Developer, QA) with a reporting chain. Every agent writes a structured report when done. Tech Lead aggregates into feature reports. Reports stored in date-organized folders. A /build skill kicks off the pipeline.

Full design: agency-company-structure


Open Questions

1. Autonomy Level

Question: Should the Tech Lead auto-pick features from planned/, or always wait for assignment?

OptionDescription
A) Assignment onlyYou explicitly say “build FR-072”. Safest, full control
B) Auto-pick with confirmationTech Lead proposes next feature, you approve before work starts
C) Fully autonomousTech Lead picks and builds. You review after

Recommendation: Option A initially, evolve to B once trust is established.

Decision:

2. Git Branching

Question: One branch per feature or per task?

OptionDescription
A) Per featurefeature/FR-072 — simpler, one PR per feature
B) Per taskfeature/FR-072/01-extract-middleware — more granular, more PRs

Recommendation: Option A — per-feature branches. Simpler. Tasks are commits within the branch.

Decision:

3. Escalation Policy

Question: When should agents escalate to you vs. decide themselves?

OptionDescription
A) Escalate architectural decisionsAgents decide implementation details, escalate design choices
B) Escalate everything unclearConservative — more interruptions, more control
C) Never escalateFully autonomous — report after the fact

Recommendation: Option A — architectural decisions and ambiguous requirements escalate. Implementation details are the agent’s call.

Decision:


Phase Overview

PhaseDescriptionStatus
Phase 1Agent definitions + report templates
Phase 2Reporting infrastructure + storage
Phase 3Orchestration pipeline + /build skill
Phase 4Daily digest (PM agent)

Phase 1: Agent Definitions & Templates —

Goal: Define the three core agent roles and report templates.

File / FeatureDetailsOwnerStatus
.claude/agents/tech-lead.mdOrchestrator agent definitionopus
.claude/agents/developer.mdImplementation agent definitionopus
.claude/agents/qa-reviewer.mdCode review agent definitionopus
vault/_templates/task-report.mdDev task report templateopus
vault/_templates/review-report.mdQA review report templateopus
vault/_templates/feature-report.mdTech Lead summary templateopus

Phase 2: Reporting Infrastructure —

Goal: Set up report storage and establish the reporting chain.

File / FeatureDetailsOwnerStatus
vault/00_system/logs/agent-reports/Date-organized report storageopus
Agent promptsAgents write reports as part of their workflowopus
Tech Lead aggregationTech Lead reads sub-agent reports, writes feature reportopus

Phase 3: Orchestration Pipeline —

Goal: End-to-end pipeline: assign feature → agents work → reports produced.

File / FeatureDetailsOwnerStatus
.claude/skills/build/SKILL.md/build FR-XXX kicks off pipelineopus
Tech Lead pipeline logicBreak down → assign → review → aggregateopus
Escalation handlingTech Lead knows when to ask managementopus

Phase 4: Daily Digest —

Goal: Automated daily summary of all agent work. Feeds into FR-037 (Daily Briefings) as dedicated sections.

File / FeatureDetailsOwnerStatus
.claude/agents/pm.mdPM agent scans reports, writes digestopus
FR-037 agent-reports-agentSection agent that reads agent reports and summarizes per-role activityopus
FR-037 cron-tasks-agentSection agent that reads cron/scheduler logs and summarizes system-triggered tasksopus
vault/30_daily/briefings/Daily digest output location (via FR-037 compiler)opus

Prerequisites / Gap Analysis

Requirements

RequirementDescription
REQ-1Custom agents infrastructure (FR-043)
REQ-2Python project scaffold for any tooling (FR-009)
REQ-3Agent git workflow for branching/PRs (FR-058)

Current State

ComponentStatusDetails
Agent definitionsNo agents built yet
Report templatesDon’t exist yet
/build skillDoesn’t exist
src/ codebaseNot started

Gap (What’s missing?)

GapEffortBlocker?
Agent definitions can be built now (plain markdown)LowNo
Report templates can be built nowLowNo
Orchestration needs FR-043 agent infrastructureMedSoft — can prototype with Task tool
Git workflow needs FR-058MedYes for Phase 3

Test

Manual tests

TestExpectedActualLast
pending-

AI-verified tests

ScenarioExpected behaviorVerification method

E2E tests

ScenarioAssertion

Integration tests

ComponentCoverage

Unit tests

ComponentTestsCoverage

History

DateEventDetails
2026-03-13CreatedInitial FR based on user idea. Design draft: agency-company-structure

References

  • Design: agency-company-structure
  • Related: FR-043 (Custom Agents), FR-056 (Autonomous Coding Orchestrator), FR-057 (Code Review Pipeline), FR-058 (Agent Git Workflow), FR-059 (Escalation Policy), FR-037 (Daily Briefings — agent reports + cron tasks feed into daily)