A report, an analysis, and an infographic on "AI agent management." It sounds like a brand-new discipline. But once you see what these texts really are, everything gets simpler — and far easier to put into practice.
Read them again in order and notice the movement: each text peels back the one before it.
Tone of "a new logic of work" and "competitive advantage." It's the pitch that sells the idea: big, inspiring, and vague the moment you try to apply it.
Separates the institutional, promotional layer from what's technically real — and admits it: this isn't any company's discovery, it's a consolidation of practices that already exist.
Distills what's left into 7 teachable items and a circular infographic. This is the part you can actually use tomorrow.
Swap the word "agent" for "new hire" and read the 7 principles again. None of them will sound strange to anyone who has ever led a team.
| Principle | In people management, this is called… | What you already do |
|---|---|---|
| 1 · Intent | Job description and objective | Explaining why the role exists and what's expected |
| 2 · Context | Onboarding and the handbook | What you teach on day one, and what you save for later |
| 3 · Data | The official system | "What counts is what's in the system, not what's on paper" |
| 4 · Success criteria | Goals and KPIs | Agreeing on what a job well done looks like |
| 5 · Autonomy with limits | Authority level and probation period | "Up to R$ 500 you decide; above that, talk to me" |
| 6 · Observation | Reporting and follow-up | Asking for the daily or weekly wrap-up |
| 7 · Supervision | 1:1 meeting and review | Reviewing, correcting, promoting (or not) |
The 7 principles describe, almost point for point, how people already work today with coding agents like Claude Code and Codex. The report translated into management language a practice developers already use.
| Principle | With a coding agent | The lesson that applies to everyone |
|---|---|---|
| 1 · Intent | The request + plan mode: the agent proposes the path before touching anything | Agree on the goal before execution |
| 2 · Context | Project instruction files (CLAUDE.md, AGENTS.md) and memory. This discipline has a name: context engineering | A short document with the house rules beats a thousand messages |
| 3 · Data | The repository is the source of truth: the agent reads the file before claiming anything about it | "Check the source, not your memory" |
| 4 · Criteria | Tests and builds: it's only "done" when verification passes, not when the agent says it's finished | Done is what's proven, not what's declared |
| 5 · Autonomy | Permission modes: plan only, ask before every action, accept edits on its own, or allow everything | It's the same N0–N4 ladder, just with a switch |
| 6 · Observation | Session history, logs, and a failure log (one line per error: what broke and the smallest fix) | A logged mistake becomes a rule; a forgotten one repeats |
| 7 · Supervision | Code review, approval before publishing, hooks that block dangerous actions | The human approves whatever is irreversible |
The 7 questions are the same ones a good manager always asked. The difference is that, with a person, a lack of clarity costs you slowly. With an agent, it costs you instantly, and at scale.
The agent doesn't fail because it's dumb. It fails because the delegation was sloppy.
It does the wrong task, very efficiently.
It picks which information to trust on its own, and picks badly.
It says "done," and you find out later it wasn't.
The infographic arranges the 7 principles around a center, not on a ladder. It's not a checklist you fill out once and forget: supervision (7) feeds back into intent (1). Every observed mistake becomes a new rule; every mistake-free week becomes one more step of autonomy.
Stop studying the theory. Answer the 7 questions and delegate your first task today.