Workflow Restructuring Feynman Cheat Sheet: Problem → Agent → Artifact — Measure Before Removing Humans from the Pipeline

Forum topic · ✨步子哥 · 2026-07-12

Summary

This Chinese forum post presents a Feynman-style cheat sheet on restructuring software workflows around AI agents, contrasting the old pipeline (problem → human → Agent → human → CI/CD → artifact) with the target state (problem → Agent → artifact). It analyzes two dimensions: organizational intelligence (upstream de-humanization via virtual PM/architect/developer/reviewer agents, with requirement understanding as the core bottleneck) and agent capability (downstream de-humanization by internalizing verification, testing, and deployment). It raises four critical counterpoints: (1) agents are non-deterministic execution engines, unlike deterministic compilers, so "requirements as source code" only holds if specs are formalized; (2) the verification gap is the true bottleneck—generation is easy, proving correctness is an order of magnitude harder, and agents reward-hack by passing tests rather than satisfying intent; (3) humans don't leave, they condense at endpoints with maximum decision density and responsibility while having less information; (4) cargo-cult multi-agent prompting ("you are a PM") doesn't create real project management. The post's central thesis: whether the restructuring works depends not on agent intelligence but on how far verification can be machine-automated, and it urges small-scale experiments measuring artifact quality before removing humans from the loop.

Workflow Restructuring · Feynman Cheat Sheet

"Problem → Agent → Artifact" — Measure Before You Pull Humans Out of the Pipeline

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

Dimension 1 · Organizational Intelligence (Upstream De-Humanization)

What humans used to do: requirement clarification, task decomposition, coordination, context handoff
New form: a virtual agent organization (PM / architect / developer / reviewer agents)
Information flow: meeting documents → communication protocols + shared context (the rationale behind MCP, A2A)
Core bottleneck: requirement understanding — agents must know when to ask follow-up questions and make reasonable 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 consistency is a hard capability)
CI/CD downgrades from a "gate" to "a function call in the agent's hands"

Hammer 1 · A Compiler Doesn't Compile While Dreaming

Assembly → C was a deterministic leap; your execution engine is an agent, which is non-deterministic
Same spec: React today, Vue tomorrow, bizarre code the day after
"Requirements as source code" ⚠️ holds only when the spec is formalized
The day it's fully formalized, it becomes a new language — you've merely changed the interface shape of programming

Hammer 2 · The Verification Gap (The Real Bottleneck)

Generation is easy; proving correctness is an order of magnitude harder (mathematical verification consumed academia for 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 systems must evolve adversarially and continuously — the hardest part to automate

Hammer 3 · Humans Don't Leave, They Condense (Responsibility Concentration)

Humans retreat to the endpoints = highest decision density + heaviest responsibility
Defining the spec is where all value judgments happen; sign-off on acceptance is the legal terminus
Whoever accepts a black-box artifact has less information but greater responsibility — more dangerous than signing off at every stage
The exception handler receives a monster that has drifted for ninety-nine steps — can they even understand it?

Hammer 4 · Cargo-Cult Multi-Agent (Watch Behavior)

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

The Only Gate

Whether the restructuring holds does not depend on how smart agents are, but on the degree to which verification can be machine-automated. Everything else is secondary.

The Bet

> Agent 60% reliable × thousands of runs × verification catches errors > humans at 99% but one item per week > > ⚠️ If verification cannot catch errors, the speed advantage inverts into disaster: sixty wrong things produced per minute.

Experimental Verification Checklist (Try on a Small Project Before You Brag)

1. Is there a real production project where quality held and delivery sped up after removing humans? (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% of the way; where taste and common sense are needed, humans return — and are more exhausted

Closing Iron Rule

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

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

---

*Feynman perspective · deconstruction of the workflow-restructuring epiphany. The test of true understanding: can you explain it to a college freshman without jargon?*