Skip to content

Notes — Blog

How the work is actually split.

A workflow, an agent, the boundary around that agent, where the cost sits, when a model is trained for the job, and a written spec. Each note says what a client gains.

01 — How the work is split

An AI workflow and an agent do different jobs.

A workflow is a path you can audit. An agent is given a goal and a boundary, and it chooses the next action. A client gains speed when each is used for the work it fits.

14 August 2026 · 5 min

Read the note

01 — AI workflow

One path. The next step starts when the last one finishes.

  1. Intake
  2. Extract
  3. Check
  4. Record

02 — Agent

A goal and a boundary. The agent chooses the next action.

Agent

  1. ReadThe source
  2. DraftThe change
  3. CheckThe rule

Goal met

A workflow is the path. An agent chooses the next action.

02 — Guardrails and governance

An agent needs a boundary, and a person who owns it.

A guardrail is the limit the agent cannot cross. Governance is the record of what it did, and the person who accepts the outcome.

28 August 2026 · 5 min

Read the note

01 — The boundary

Written before the run. A prompt does not widen it.

  1. Tools

    Which calls are allowed

  2. Data

    Which sources it may read

  3. Actions

    What it may change

  4. Stop

    When the run ends

02 — The record

Each action is visible. A person accepts the outcome.

  1. ReadThe source
  2. DraftThe change
  3. HeldPast the boundary

A person accepts

The agent acts inside the boundary. A person accepts the outcome.

03 — Where the cost sits

Pay for the step that needs a model.

Most of the spend is the wrong shape of work, a context that is too large, and a run that is allowed to continue after it should have stopped.

11 September 2026 · 4 min

Read the note

01 — The whole context

The repository, the policy, and the thread, on every call.

Paid again on the retry

02 — The step

The expensive call is the one the spec says is hard.

  1. ExtractA small model
  2. JudgeThe hard step
  3. CheckA rule
Pay for the step that needs a model. Stop the run when the goal is met.

04 — Custom models

Train a model for the job, not for a demo.

A custom model earns its place when the same step keeps missing, the examples already exist, and a person can tell whether it works.

25 September 2026 · 5 min

Read the note

01 — The path

The job is written first. The model is adapted to the examples.

  1. JobWhat must be true
  2. ExamplesFrom the real work
  3. AdaptA model that exists
  4. ProveCases it has not seen
  5. AcceptA person

02 — Held back

These cases judge the model. It does not see them while it trains.

  • The clause it must keep
  • The field it must read
  • The decision it must leave

Held-out cases pass

Train for the job. Prove it on cases the model has not seen.

05 — Spec-driven development

Write the spec before the model writes the code.

The spec says what will be true, what must not change, and how anyone can tell. AI drafts and tests against that document. A person still accepts the result.

9 October 2026 · 4 min

Read the note

01 — The spec

Written first. The pod is accountable to it.

  1. Outcome

    What will be true

  2. Constraint

    What must not change

  3. Done when

    How anyone can tell

02 — Checked against it

The draft, the tests, and the acceptance follow the spec.

  1. BuildAgainst the clauses
  2. ProveCases from the spec
  3. AcceptA person

Meets the spec

The spec is written first. Build and proof are checked against it.

06 — AI in engineering

Use AI to engineer the software. Keep a person on the outcome.

Assisted development, intelligent testing, and automated engineering workflows shorten the path from a written outcome to a release. Accountability for that outcome stays with a person.

23 October 2026 · 4 min

Read the note

In sequence

  1. Write
  2. Test
  3. Review

Test runs with the write

  1. Write
  2. Test
  3. Review
Same three steps. The test agent overlaps the write, so review starts sooner.

07 — AI in business

Put the assistant inside the process.

An agent, an enterprise assistant, or a document workflow earns its place when it takes a step the business already runs, and leaves a record of what it did.

4 November 2026 · 4 min

Read the note

01 — Document

Invoice, contract, or claim.

02 — Fields

  • PartyAcme
  • Amount4,200
  • Date08 Nov

03 — Decision

  • MatchedContinues
  • ExceptionA person
The file becomes fields, then a decision. The exception waits for a person.

Start a brief

Bring the outcome.

Say what should be true when the work is done. We will say whether it is a workflow, an agent, or both, and which offer it belongs to.

Start a briefRead the model