agent-delegation

Agent Delegation

Overview

Determines which agent runtime executes a given story, with what constraints, and under what governance. Prevents misassignment incidents (e.g., Copilot modifying existing files) and maximises parallel execution throughput.

When to Use

• Sprint planning: assigning stories to runtimes
• Deciding whether a story needs Claude Code, Copilot, sub-agents, or human
• Launching parallel sub-agents for independent work
• Reviewing whether an assignment follows governance rules

When NOT to Use

• For sprint ceremony process — that's sprint-management
• For ADR creation and review — that's adr-governance
• For code quality standards — that's python-quality

Core Process

Agent Runtimes

Runtime	Role	Strengths	Constraints
Claude Code	Orchestrator	Full context, multi-file edits, MCP tools, Playwright, sub-agents	—
Claude Code Sub-Agents	Parallel workers	Concurrent execution, isolated context	Cannot share state or modify same files
GitHub Copilot	Autonomous PR creator	Creates PRs with CI, works async	DESTROYS existing files. Limited context. No MCP.
Human	Governance	Credentials, approval, strategic judgment	—

Assignment Rules (first match wins)

Rule 1: Credentials Required
  → HUMAN (Claude Code assists via Playwright)

Rule 2: Modifies Existing Files
  → CLAUDE CODE (NEVER Copilot)

Rule 3: Multiple Independent New Files
  → CLAUDE CODE SUB-AGENTS (parallel)

Rule 4: Single New File with Dependencies
  → CLAUDE CODE (sequential)

Rule 5: Research or Exploration
  → CLAUDE CODE SUB-AGENT (Explore type)

Rule 6: Architecture Decision
  → CLAUDE CODE + Plan agent, REQUIRES human approval

Governance Gates

Gate	Trigger	Approver
Sprint planning	Before any story assignment	Human
Architecture review	ADRs, new agent types, schema changes	Human
PR review	Before merge to main	Human
Skill modification	Any change to skills/*/SKILL.md	Human (must cite data)
Credential operations	New API keys, service accounts	Human executes directly
Destructive git ops	Force push, reset, branch delete	Human confirms explicitly

Common Rationalizations

Rationalization	Reality
"Copilot can handle this small edit to an existing file"	Copilot rewrites entire files, losing existing logic and structure — assign to Claude Code
"We'll run these stories sequentially, it's simpler"	Independent stories executed sequentially take 2-3x longer; sub-agents run them in parallel
"We don't need human approval for this architecture change"	Agents making structural decisions without governance creates direction conflicts — ADRs need sign-off
"Let's just do it without a GitHub issue"	Untraceable work products break sprint discipline and audit trail
"Sub-agents can coordinate on shared files"	Sub-agents have isolated context; sharing state between them causes race conditions and overwrites

Red Flags

• Copilot assigned to a story that modifies existing files
• Independent stories executed sequentially instead of in parallel
• Architecture decision made without human approval gate
• Work started without a GitHub issue
• Sub-agents assigned to modify the same file
• Credential operation attempted by an agent instead of human
• Story assignment doesn't match any rule in the assignment matrix

Verification

• Every story has a designated runtime before implementation begins — evidence: sprint plan includes runtime column
• No Copilot assignments touch existing files — evidence: Copilot stories only create new files
• Independent stories run in parallel where possible — evidence: sub-agent logs show concurrent execution
• Architecture decisions have human approval — evidence: GitHub issue comment from human before implementation
• All work has a GitHub issue — evidence: every PR references an issue number
• Governance gates enforced — evidence: PR review required, no force pushes without explicit approval