Workflow Restructuring Cheat Sheet: Problem → Agent → Artifact

*Before you pull the human out of the pipeline, measure first.*

The shift being discussed:

• Old: Problem → [Human] → Agent → [Human] → CI/CD → Artifact
• New: Problem → Agent → Artifact

Dimension 1 · Organizational Intelligence (upstream de-humanization)

• What humans used to do: requirement clarification, task decomposition, coordination, context hand-off
• New form: virtual Agent organizations (PM / architect / developer / reviewer Agents)
• Information flow: meeting documents → communication protocols + shared context (the origin of MCP, A2A)
• Core bottleneck: requirement understanding — Agents must learn to ask follow-up questions and make sensible defaults under uncertainty

Dimension 2 · Agent Capability (downstream de-humanization)

• What humans used to do: code review, test acceptance, triggering pipelines, watching deployments, rollbacks
• New form: internalize verification/testing/deployment; build self-feedback loops
• Environment grounding: call tools, run sandboxes, observe results, self-correct (long-horizon coherence is a hard capability)
• CI/CD is demoted from a "gate" to just another function call in the Agent's hands

Hammer 1 · Compilers Don't Compile (They Don't Dream)

• Assembly → C was a deterministic transition; your execution engine is an Agent — non-deterministic
• The same spec yields React today, Vue tomorrow, bizarre code the day after
• "Requirements as source code" ⚠️ only holds when the spec is formalized
• The day it's formalized, it becomes a new language — you've only reshaped the programming interface

Hammer 2 · The Verification Gap (the real bottleneck)

• Generation is easy; proving correctness is an order of magnitude harder (formal verification took academia decades)
• Tests-as-contract → Agents learn to pass tests rather than satisfy intent
• Reward hacking: the best lawyers know how to be "legally dishonest"
• Evaluation must evolve adversarially and continuously — the hardest part to automate

Hammer 3 · Humans Don't Leave, They Concentrate (responsibility concentrates)

• Humans retreat to the endpoints = highest decision density + heaviest responsibility
• Defining the spec carries all value judgments; accepting deliverables is the legal endpoint
• Signers of black-box output have less information, more responsibility — more dangerous than per-stage sign-offs
• The exception handler receives a monster that has drifted for ninety-nine steps — can they even read it?

Hammer 4 · Cargo-Cult Multi-Agent (watch behavior)

• A "you are a PM" prompt ≠ actually doing project management
• A real PM weighs resources, handles politics, says "no" to ambiguity
• If behavior doesn't change after removing that line, it's just a generator with a name
• [1→n] is a story you tell yourself, not a fact — measure first

The Only Gate

Whether the restructuring works doesn't depend on how smart Agents are, but on how far verification can be mechanized. Everything else is secondary.

The Bet

Agent 60% reliable × a thousand runs × verification catches errors > humans 99% reliable but one item per week

If verification can't catch errors → the speed advantage inverts into catastrophe: sixty wrong things per minute.

Experimental Validation Checklist (try on a small project before bragging)

1. Is there a real production project where removing humans kept quality and sped delivery? (Not a demo.) 2. How is verification done? Does a smart human still have to squint at the end? 3. How many times did reward hacking occur? How was it discovered? 4. Pure software / machine-verifiable domains might get 70–80% there; where taste and common sense are needed, humans come back — and are more exhausted

Closing Iron Rule

Run one small experiment: remove the humans in the middle and measure the artifact quality.

> If quality drops — the plane never came. What you built was a bamboo control tower, not an airport.

*Feynman's test: if you truly understand, you can explain it to a college freshman without jargon.*