AI agent management · straight to the point
It's not enough to create agents. You have to manage them.
Everything you need to understand about the new way of working with AI, in 7 principles, plus a tool that applies all 7 in 5 minutes.
Everything you need to understand about the new way of working with AI, in 7 principles, plus a tool that applies all 7 in 5 minutes.
Before, you used AI to answer questions. Now it can interpret what you want, look up information, execute tasks and deliver a result, with more or less freedom. That changes your role.
Instead of configuring every step, you define the goal, what counts as a good result, the limits and when it's done. AI decides part of the path.
The agent is only as good as the information it uses. Someone has to say which source is official and who wins when two sources disagree.
You move from "doing the task" to "managing the system that does the task": vision, context, judgment, evaluation and correction.
Each principle answers a question you'd ask a new employee. The Agent Sheet asks exactly these questions for you.
The agent needs to know the goal, why it matters and what result is expected. Without that, it does the wrong task efficiently.
Rules, examples, history, current state. But more context isn't better: the right amount is the minimum that produces the best decision.
Official source, consistency, freshness, traceability. Wrong data in, wrong decision out.
Metric, expected quality, acceptance condition, definition of "done". Without it, it finishes without solving anything.
Not every agent deserves the same freedom. It grows in levels, from N0 to N4, together with trust.
Record results, errors, costs, time and decisions. Execution generates evidence; evidence generates learning.
AI executes within limits, but responsibility stays yours: goals, exceptions, corrections, evolution.
What you learn in supervision adjusts intent, context and rules. That's how isolated agents become a working system.
The more freedom, the more control. The agent starts low and climbs when it shows, with evidence, that it deserves to.
# The rule the Agent Sheet uses to calculate your level Involves money, an irreversible action, or outside people? → at most N2 (you approve) Internal and reversible? → can reach N3 Only looks up information or drafts? → N0 / N1 No official source defined? → stays at N1, no exceptions
Reports about "the new era of agents" come loaded with promotional layers. This isn't an exclusive discovery of any one company: it's the consolidation of what already works in practice. After the filter, 5 truths remain.
The agent finds conflicts and duplicates, but the organization decides which piece of information has authority.
You provide goal, context, constraints, criteria and expected result. The system chooses part of the path.
It's not enough to say what to do. You need to say what's good, how to measure it and when it's actually done.
Permissions, autonomy levels, human review, monitoring, policies and security. More autonomy, more control.
The hard part is no longer executing. It's defining intent well, building context, organizing data, evaluating and correcting.
Working in a demo proves nothing. In real use, what matters is stability, repeatability, safety and how it handles the unexpected.
The new discipline isn't "using AI."
It's AI Agent Management.
The Agent Sheet is a free page that runs in your browser. Nothing is sent anywhere. You answer, it builds.
Click a ready example (clinic, store, accounting office, teacher) and adjust it to your case, or start blank.
Open the Agent SheetWrite the way you'd talk to a person. Every field has an example. On question 5, check whether the task involves money, an irreversible action, or outside people.
There are 4 tabs: the Sheet (a summary of your agent's 7 principles), the Ready instruction (to paste into ChatGPT, Claude or Gemini), the 3 tests and the supervision Checklist. At the top, the calculated autonomy level and what's needed to level up.
# excerpt from a generated instruction
## Your autonomy level: N2 (Prepares, you approve)
You prepare everything, but NOTHING is sent, changed or executed
without my explicit approval. Always show me first and wait for my "OK".
Run the 3 tests: missing information, forbidden request and alarm situation. Did it fail one? Adjust the sheet. Want to polish it? The ✨ Refine with AI button copies a request for ChatGPT/Claude to review your sheet.
2 minutes a day reading the report, 10 minutes a week adjusting one rule. Autonomy only grows with evidence. When in doubt, drop a level.
30% of patients missed appointments without notice. In 5 minutes with the Sheet, she came out with this:
Confirm next day's appointments via WhatsApp, to reduce no-shows.
The clinic's scheduling system. Spreadsheets don't count. Something odd: it flags it, never guesses.
Because it talks to patients and a sent message can't be undone. Forbidden: rescheduling, cancelling, discussing prices. Moves to N3 after 4 error-free weeks.
At 5pm: sent, confirmed, no response, errors. Calls right away if there's a complaint. Review every Friday, 10 minutes.
There's a reading behind the report, the analysis and the infographic, and it changes how you see all of this. Spoiler: you probably already practice at least 5 of the 7 principles, just under a different name.
The original material is in the repository: report, analysis and the 7 principles.
/ficha-agente skill10 fields, from N0 to N4, ready to open and adjust. And, for Claude Code users, a skill that runs the interview, writes AGENTE.md and audits an agent you already have.