Skip to main content

Implementation Agent

Position in pipeline: Agent 6
Phase: 4 — Agentic Implementation
Min confidence: 0.75
Version: 1.0.0

Role​

The Implementation Agent is the orchestrator of code generation. It does not write code directly in one context — it spawns context-fresh sub-agents for each task in the execution wave plan, ensuring that each sub-agent operates with a clean, minimal context window.

This is the highest-throughput agent in the pipeline. A squad's Implementation Agent can execute multiple features in parallel waves while humans review confidence scores and resolve escalations.

The orchestrator-executor pattern​

The Orchestrator:

  • Reads tasks.md and manages wave sequencing
  • Spawns context-fresh sub-agents for each task
  • Collects outputs and confidence scores
  • Calculates per-wave CCS
  • Escalates to the Knowledge Agent when wave CCS falls below threshold

What each sub-agent receives​

Each context-fresh sub-agent receives a minimal context payload constructed by the Knowledge Agent:

task: TASK-001
requirement: REQ-001 — "When authenticated merchant submits payment, the system shall..."
local_files:
  - src/domain/Payment.ts
  - src/domain/types.ts
domain_terms:
  - term: Payment
    definition: "..."
  - term: Settlement
    definition: "..."
steering_rules:
  - "Business logic must reside in service classes, not controllers"
  - "Controllers may not access database repositories directly"

Nothing from other tasks, other slices, or unrelated domain areas is included.

TDD workflow​

The Implementation Agent follows a strict test-driven sequence for every task:

RED    → Sub-agent writes a failing test mapped to REQ-NNN acceptance criteria
GREEN  → Sub-agent implements logic sufficient to pass the test
REFACTOR → Refactor Agent improves code quality within spec boundaries

Every generated test file is named after the requirement it satisfies: REQ-001.spec.ts

Undocumented behavior detection​

Code that introduces behavior not defined in any spec requirement is flagged as UNDOCUMENTED_BEHAVIOR and blocked from merge:

[UNDOCUMENTED_BEHAVIOR] src/domain/Payment.ts:47
  Method applyDiscount() is not referenced in any spec requirement.
  Add to requirements.md or remove from implementation.

Goal-backward verification​

After completing each task, the sub-agent performs Goal-Backward Verification:

  1. Re-reads the original REQ-NNN requirement
  2. Verifies that the implementation satisfies the acceptance criteria
  3. Confirms no behavior outside the requirement was introduced

This prevents "technically complete but spec-noncompliant" outputs.

Governance​

RuleBehavior
Confidence < 0.75Commit to feature branch only; mandatory human code review
CCS below 0.65Wave halted; all downstream tasks paused; TL escalation
Undocumented behavior detectedBlocked from merge; TL notified
No requirement referenceTask rejected; returned to Task Planning Agent

Human code review trigger​

Human code review is mandatory when:

  • Implementation Agent confidence < 0.75 for any task in the wave
  • The wave CCS falls below 0.65
  • The implementation touches auth, payment, or data-access code

Next agents​

  • QA Agent — validates spec coverage and test coverage
  • Refactor Agent — improves code quality within spec boundaries (runs within Phase 4)