Patterns for reliable reasoning

Problem. Decisions are forgotten, rationalised after the fact, and never learned from.

For people

Before deciding, write down the intent, options, expected outcome, and your confidence. Review it when the outcome is known.

For agents

Log intent → policy check → approval → action → evidence → outcome as one durable record per consequential action.

Failure mode
Logging only the final answer, so nobody can reconstruct why it happened.

Problem. Plans look safe because nobody is asked how they could fail.

For people

Imagine it is a year from now and the plan failed. List the most likely reasons, then mitigate the top three.

For agents

Before an irreversible action, have the planner list failure scenarios and check each against deterministic policy.

Failure mode
Running the premortem with the same blind spots, e.g. the same model judging its own plan.

Problem. Whoever proposes an action also gets to approve it.

For people

Use a cooling-off rule: big decisions are proposed one day and confirmed the next, ideally by a second person.

For agents

The model proposes; an independent gateway authorises. Prompts such as "do not spend money" are not a control.

Failure mode
A supervisor that rubber-stamps because it shares the proposer’s incentives or model.

Problem. An approval is given for one thing and used for something slightly different.

For people

Approve the exact document version, not "the contract". Any edit means approving again.

For agents

Bind each human approval to a hash of the exact action payload, with an expiry and single use; re-check it at execution time.

Failure mode
Approving "send the email", then the agent edits the recipient list.

Problem. "Done" is reported without proof that anything actually changed.

For people

Close a task only when you can point to the artifact: the receipt, the merged PR, the signed form.

For agents

An HTTP 200 is not success. Read back the external state and store the evidence before marking the goal complete.

Failure mode
An agent confidently reports a payment, order, or account that never existed.

Problem. A single typo or misreading produces an absurd outcome.

For people

Define normal ranges in advance (budget, quantity, time) and pause when you are outside them.

For agents

Deterministic bounds on quantities and amounts relative to history; anything outside the envelope needs approval.

Failure mode
Ordering thousands of units when the usual order is dozens.

Problem. Confidence does not track accuracy, and nobody measures it.

For people

Attach a probability to predictions and score yourself over time.

For agents

Compare stated confidence against verified outcomes; route low-calibration task types to stronger models or humans.

Failure mode
Measuring activity (messages sent, hours run) instead of verified outcomes.

Problem. Instructions and facts are accepted from anyone who sounds authoritative.

For people

Verify who is asking through a second channel before acting on unusual requests.

For agents

Tag every memory and instruction with its source and trust level; untrusted content can never grant permissions.

Failure mode
An agent accepts an unverified claim of authority, or a prompt injected into a web page.