Back to detailsPutting Agents Into Real Workflows

01 / Reframe

An agent is not a tool list

A better representation is this: an agent is a work unit placed inside context, roles, artifacts, and feedback loops.

Tools focus on capability
Workflows focus on position
Systems focus on feedback

02 / Resources

Three primary objects

A real workflow is not a chat trace. It is a set of objects that can be identified, assigned, checked, and reused; feedback makes them loop.

Context

Goals, sources, constraints, decisions, and current state.

Role

Reader, Planner, Reviewer, and Writer are work boundaries.

Artifact

A document, task, judgment, code change, or reusable record.

03 / Flow

How one collaboration flows

Work is not a single exchange but a loop you can run again: goals go in, judgment comes back.

Input

Set the goal

Name what to solve and where the edges are.

Context

Gather context

Hand the agent the material, limits, and past decisions.

Run

Run and record

Run the tools and produce a checkable artifact.

Feedback

Write judgment back

Human trade-offs become the next round's context.

04 / Baseline

Get to a baseline

Do not optimize the prompt first. Make the workflow repeatable, recordable, and comparable.

workflow-baseline — zsh

git clone <repo> && cd agent-workflowpnpm installcp .env.example .envpnpm eval --case beforepnpm run session -- --record

Every improvement needs the same inputs and the same judgment surface.

05 / Evals

Define quality with verifiable tasks

Evaluation is not a school test. It tells the workflow which part is actually improving.

ID	Task	What it tests	Grader
R1	Extract context	Missing constraints	human spot-check
R2	Generate plan	Executable steps	structure match
R3	Draft artifact	Fit for audience	human score
R4	Review risk	Key assumptions	checklist
R5	Handoff summary	Reusable next time	reuse rate

Score = repeatable input + inspectable artifact + human judgment.

06 / Dashboard

Know whether the workflow got better

A strong agent system is not more fluent. It has shorter context, steadier judgment, and more reusable artifacts.

4work objects

3checkpoints

0hidden steps

82%reuse rate

Context

Shorter input

Background compresses into reusable chunks.

Artifact

Steadier output

Every run leaves something inspectable.

Feedback

Judgment returns

Human tradeoffs enter the next round.

07 / Doer vs Tutor

The Doer / Tutor boundary

The same agent behavior can help you form a mental model, or let you bypass one.

Doer

Hands over the answer

It removes search, but also removes understanding, tradeoffs, and internalization.

Tutor

Provides scaffolding

It lowers extraneous load while leaving the key judgment to you.

08 / Feedback

Write feedback into the next run

A workflow improves because human judgment becomes future context, not because the agent remembers more.

feedback-loop.ts

artifact = agent.run(context)review = human.check(artifact)memory.add(review.decision)rules.add(review.risk)next.run(memory, rules)

Record decisions

Why this tradeoff was chosen.

Keep risks

What could invalidate the result.

Compress rules

What should be reused next time.

09 / Shift

From tool list to work system

Same AI, a different representation — and a very different long-term result.

Before

Tool list

Models, plugins, commands, and buttons; it scatters as you learn.

After

Work system

Objects, mechanisms, feedback, and delivery; it clarifies as you use it.

10 / Inner Map

What remains is an inner map

A good agent workflow does not only finish the task. It helps you recognize the next situation faster, judge the path better, and ship more steadily.

Redraw the object: what am I operating on
Compress the structure: can it work next time
Keep the judgment: who owns the tradeoff