Skip to content

The Human Layer

Agents execute. Humans decide.

Why Humans Are the Critical Path

Project Phoenix is a six-agent pipeline. But the seventh participant — the human — is the most important one. AI agents can extract business logic, map interfaces, synthesize requirements, design architectures, build code, and validate results. What they cannot do:

  • Interview stakeholders who carry tribal knowledge
  • Resolve ambiguity when the code contradicts the business intent
  • Make judgment calls about which business rules are still relevant
  • Sign off on the specification before building begins
  • Decide go/no-go at every gate in the pipeline

These aren't limitations of current AI. They're structural properties of the problem. Legacy systems exist in an organizational context that no codebase can fully capture.


The AI Software Lead

In a Phoenix engagement, the human role is the AI Software Lead — a senior practitioner who:

  1. Orchestrates the six agents across the pipeline
  2. Validates outputs at every stage before the next agent begins
  3. Supplements agent findings with stakeholder interviews and domain expertise
  4. Resolves conflicts between what the code does and what the business needs
  5. Makes the architectural decisions that require judgment, not just analysis

This isn't a project manager role. It requires deep technical understanding combined with business acumen — the ability to read an agent's business rules catalog and know whether it captured the right intent.


Human Touchpoints in the Pipeline

Pipeline StageHuman Responsibility
Agent 1: ExtractorValidate business rules against stakeholder knowledge. Flag rules that are obsolete vs. essential.
Agent 2: ArchaeologistConfirm user journeys match actual workflows. Identify undocumented processes.
Agent 3: SynthesizerReview unified spec for completeness. Resolve ambiguity flags raised by the agent.
Agent 4: ArchitectApprove tech stack and architecture decisions. Confirm alignment with organizational constraints.
Agent 5: BuilderReview critical code paths. Validate implementation matches specification.
Agent 6: ValidatorFinal sign-off on regression results. Accept or reject the gap exceptions list.

What Agents Can't Know

Every legacy system contains knowledge that lives outside the code:

  • Tribal knowledge — "We do it this way because of a regulatory change in 2012 that nobody documented"
  • Political context — "That module exists because two departments couldn't agree on a shared process"
  • Future intent — "We've been meaning to change that workflow but never had the budget"
  • Edge cases from experience — "That validation rule exists because of a specific incident with a specific client"

The AI Software Lead bridges the gap between what the agents extract from code and what the organization actually needs.


The Gate Model

Phoenix operates on a gate model — no agent's output passes to the next stage without human review and approval.

[Agent 1] → GATE → [Agent 2] → GATE → [Agent 3] → GATE

[Agent 6] ← GATE ← [Agent 5] ← GATE ← [Agent 4] ← GATE

At each gate, the AI Software Lead can:

  • Approve — output is complete and accurate, proceed
  • Supplement — output needs additional context from stakeholders
  • Revise — agent missed something, re-run with additional guidance
  • Escalate — fundamental ambiguity requires business owner decision

This ensures that no compounding errors propagate through the pipeline. A mistake caught at Agent 2 doesn't become an architectural flaw at Agent 4.


"The agents see the code. The human sees the context."

The Economics →