Working inside the loop · five sections

THE
AGENTIC
PDLC

An agent is a colleague, not a tool. The whole of this page follows from that: if you hand a colleague a finished answer and ask them to type it up, you have hired an expensive typist. The work is to give the agent the problem, leave it room to decide, and then judge what comes back.

The core principle

"Every decision you make for the agent before it sees the problem is a decision you paid human rates for that it could have made in seconds."

This came out of moving a production engineering organization onto agents. Almost every way a team leaves value on the table turns out to be a version of deciding too early.

01 The thesis

The Value Curve

Value scales with decision space. The more decisions the agent makes on your behalf, the more you get out of it. Hand over a problem and it explores the solution space and makes hundreds of calls. Hand over a finished mock and it makes almost none.

Max leverage Expensive typist
Problemagent explores; hundreds of decisions Constrained problembounded, still 50+ real decisions Functional specimplementation calls only Mock translationintent to code; no decisions
02 Structure

The Seams Principle

Every spec you write creates a seam, and your team owns every seam. One spec covering a whole feature has none: the agent holds the connections itself. Split that feature five ways and you have bought four integration problems.

One spec Complete feature Zero internal seams. The agent manages every connection.
Five specs ×××× Four seams, all yours: integration, coherence, debugging.
For practitioners: the decision test
The test

If you changed the core entity's schema, how many specs would break? If the answer is more than one, they should have been one spec.

Core entityWhat belongs in one spec
Support cases Intake, triage, assignment, communication, resolution, reporting
Bulk uploads Upload, validation, processing, error reporting, history
Appointments Scheduling, queue management, resource matching, notifications
Load-bearing principles
  • Split by domain, not by team. Different users looking at the same system is one spec. Different systems sharing a user is two.
  • Personas are not a seam. A customer-facing upload flow and the internal screen that manages the same records are one spec, not two, however many teams touch them.
  • Phases are seams in disguise. Plan, then design, then build, with gates between and no re-litigating, is three specs and two seams. It recreates the handoff problem in a new medium.
  • Watch for the single author. If a spec is so specific that only its writer can produce or judge it, that person is now the chokepoint for all the work.
03 The irreplaceable part

What the Human Brings

The role does not shrink. It shifts up the stack. Four things an agent cannot get on its own, and they are the four worth your morning.

The problem, and why now
Business context and customer pain. The agent has no way to know either.
Constraints as durable skills
Regulation, business rules, org politics, the mission. Encode them once where every run can read them.
Success criteria as outcomes
Not features. "No-show rates drop by half." Let the agent generate the features that hit the number.
The messy context
Customer quotes, competitor screenshots, the long Slack thread. Do not filter it. Let the agent synthesize.
The role shift
TraditionalWorking with agents
Write detailed user stories Define the problem and the success criteria
Specify UI flows Evaluate the approaches the agent proposes
Decompose into tasks Set constraints and priorities
Manage a backlog of tickets Steer direction between iterations
Write acceptance criteria Define measurable outcomes
Groom with engineering Hand over messy context and let the agent sort it
04 The rhythm

Steering, Not Driving

The cycle alternates. You set the problem, the agent explores, you react to what it actually produced, it refines, you judge against the criteria. Every human turn is a reaction to real output, never a guess about what the output might be.

  1. YouProblem, constraints, context, success criteria
  2. AgentExplores approaches, proposes solutions
  3. You"Approach two, but you missed the data-retention angle"
  4. AgentRefined spec, breakdown, implementation
  5. You"This meets the criteria," or "this misses X"
Steering"I like this direction, but account for customers with 50,000+ end users" Driving"Move the button left and make it a dropdown"
The design partnership

A designer reviewing three generated approaches and saying "that one, but fix the spacing" beats mocking one approach and hoping it survives translation. The trade is prescription for curation, and it runs on one rule about how specific to be:

Visual specificityHurts. The agent cannot see your mock, and pixel fidelity is slow to converge. Functional specificityHelps. "Validate the headers before submitting" is testable acceptance criteria. Unstated expectationsWorst. Assumptions that feel obvious produce "no, not like that" five times running.
05 Check yourself

Red Flags

Warning signs, before the next planning session. Recognize any of these and the move is the same: go back and re-engage the agent with the problem instead of the solution.

  • The spec describes screen layouts to the pixel.
  • A requirements doc exists, and it already picks phases or a solution.
  • Two or more specs share a core entity.
  • Someone spent more than a week on the idea before an agent saw it.
  • User stories describe micro-interactions instead of outcomes.
  • Design produced mocks before the agent explored any approach.
  • Specs are split along team boundaries instead of domain boundaries.
  • A prompt prescribes output format, section structure, or a deliverable template.
  • The work is split into phases with rigid handoffs between them.
  • A spec exists that only one person can write or judge.

If you checked any of these, you are paying human rates for work that runs at machine speed.

The next rung

Everything above keeps a person in the loop.

Each cycle still needs someone to aim it and someone to judge what came back. That works, and it is where most teams should be. It also has a ceiling: you are the throughput. Loop & Harness Engineering is what you build when you would rather be on the loop than in it, with the judging wired into the system and your attention on whether the judging is any good.