Skip to main content

Best practices : the discipline of reliable agents

Tools and configuration take you far ; discipline takes you the rest of the way. These practices come from months of operating agents in production at eaQbe : every one of them exists because its absence once cost us something. They are simple to state, and all of them can be enforced by configuration rather than left to good intentions.

Premortem : check before proposing​

A « premortem » is the habit of verifying before asserting. Before the agent proposes a change or states a fact, it must answer : what have I actually read or measured that backs this? No source, no claim.

Why it matters : language models are eloquent, and eloquence makes unverified statements sound true. The premortem rule converts « it sounds right » into « I checked it ». In our configuration this is a hook : when a request touches sensitive topics, an automatic reminder fires : cite what you read, re-measure every number, do not propose solutions before diagnosing.

A concrete effect : asked « what's wrong with X? », a premortem-disciplined agent answers « let me read X first » : and then reads it : instead of producing a plausible-sounding guess.

Verify before « done »​

The mirror-image rule, applied after the work : never claim a result without evidence. The test that passed, the page that responds, the number that matches the database : proof first, claim second.

This sounds obvious and is constantly violated : by people and by AI alike. The practical trick is to make verification part of the task definition : the work is not « update the page » but « update the page and show me it responds correctly ». Agents follow the task definition : so put verification inside it.

The lessons loop : every mistake happens once​

When something goes wrong, the lessons file gets a new entry : what happened, why, and the rule that prevents it. The file is re-read every session. Over months this compounds into something remarkable : a system whose error rate decreases with use, because its history is institutional, not anecdotal.

The habit to build : end every incident, however small, with the question « what rule would have prevented this? » : and write it down. Thirty seconds of writing buys you immunity to a whole category of repeat mistakes.

Structured sessions : start informed, end recorded​

Two bookends make long-running collaborations reliable :

  • Session start : the agent loads its context before working : project rules, memory, lessons, the current state of the systems it depends on. Like a pilot's pre-flight checklist : boring, mechanical, and the reason nothing is forgotten. ( Ours is automated : one command, and the agent reports « context loaded : N rules, M memories, systems green » before touching anything. )
  • Session end : significant decisions and outcomes get saved to persistent memory, so the next session : tomorrow or in three months : starts from the record.

One change at a time​

When improving any system with an agent, change one thing, measure, then change the next. Multiple simultaneous changes make results impossible to attribute : you learn nothing even when things improve. This rule is centuries old in experimental science ; agents simply make it easy to violate, because changes have become so cheap to produce.

Frequently asked questions​

What is a premortem in AI work?​

The discipline of verifying before asserting : reading the actual file, measuring the actual number : before proposing anything. It is the antidote to plausible-sounding-but-wrong output.

How do you stop an AI agent from claiming work is done when it is not?​

Put verification inside the task : « do X and show evidence it worked ». Enforce it with hooks where possible. Evidence before claims, mechanically.

Do these practices require technical skills?​

No : they are working habits, written in plain language in the agent's instruction files. What requires a bit of setup is automating them ( hooks ), which is part of what we cover in training.

Going further​

  • Final page : the eaQbe-ISO standard : everything in this guide, assembled into one ready-to-copy configuration.
Put this into practiceeaQbe, the team behind these tutorials, delivers hands-on trainings and consulting in Brussels and remotely.