Semantic levelsees the whole sea

Operational levelworks the ropes

Grasps the big picture

Reads backlog, documentation and meeting notes in one go.

Runs rule-based automation

The repetitive steps, the same way every time.

Pulls out what matters

Finds the thread in hours of discussion.

Cuts overhead

Status, summaries and docs without the busywork.

Analyses and categorises

Sorts, clusters, spots duplicates and dependencies.

Decodes black boxes

Reverse-engineers tools nobody documented.

Smooth sailing, right?

But …

Speed without a heading is drift

That deckhand belongs at the spyglass, not at the helm.
At the helm, every mistake becomes part of the system, and the next sprint builds on it.

Chart Nº 01 · the course of a crew

Planned headingDrift: small errors, left uncorrectedCompounding: the next sprint builds on the error

Drift ends in the wrong harbour. Once the mistake is part of the system, the crew ends up somewhere else entirely.

Kinds of drift

Five ways off course

The dotted line is the heading. Drift rarely announces itself.

Goal driftThe harbour moves. The goal gets reinterpreted, and nobody decided it.

Scope driftSame harbour, wider course. More modules, interfaces and data than asked for.

Solution driftThe rabbit hole. New layers and frameworks for a small request.

False greenGreen but useless. Every check passes, yet the user cannot do what was asked.

Rule driftThe rules become the goal. No merge over a bug with zero impact.

How the drift shows

Typical indicators. Not a complete list, but the ones that show up most often.

Sign 1Stories outpace their readers Sign 2Ping-pong between the coding agent and the review agent Sign 3Nobody knows who checked what Sign 4False green: every check passes, the goal slips Sign 5Too smooth: the AI agrees with everything Sign 6Nothing is reconciled (no single source of truth) Sign 7The user test is not what anyone expected Sign 8Context rot: the context window is over 65% full Sign 9The agent rereads the same files again and again

Checkpoints against drift

Drift is easier to prevent than to undo. Each checkpoint surfaces misunderstandings, including a person's own, and is a safe harbour to roll back to.

Standing orders

AI is probabilistic.
Without a deterministic system around it, drift is the course from day one.

DdeterministicPprobabilisticHhuman authority

Human gateH

Every commit gets a visual check by a person before it counts.

Second opinionP

No self-certification. An outside AI reviews what the first one built.

QA hooksD

Checks against green but useless: user value is the goal, not passing tests.

Reconcile on mergeD

After every merge: does everything still fit together?

Sceptics on boardP

Stop hooks call in sceptical sub-agents. Does this still serve the people who use it?

TriangulateP

A base prompt against sycophancy. Every claim checked from three sides: thesis, antithesis, synthesis.

Last line of defence

Merge gates and release gates stop what slipped past everything else. If drift gets through anyway:

RepairThe build is wrong, the goal holds.
Fix it in the same slice.

ReshapeThe goal or the assumption is wrong.
Back to the product decision.

RevertAlready merged.
Back to the last buoy that held, or start fresh on a feature branch.

A rollback is not a data restore.

Self-check: does your system have drift?

More on the subject …

… in the Logbook

Entry Nº 12

AI in agile teams

AI is a very self-assured deckhand.