Agent design patterns
Eight ways agent products lose trust
The failures that recur across agent products. Each one names what it costs and the interaction pattern that addresses it.
| What it looks like | What it costs | Instead | Addressed by |
|---|---|---|---|
| Confirm only at the end The agent works for twenty minutes and then asks whether it got it right | One error in step three means redoing all twelve | Place checkpoints deliberately, weighted early | Scheduled confirmation |
| Transparency theatre A wall of tool calls labelled reasoning | Overload reads as opacity and nobody checks anything | Summary by default, detail on request | Altitude control |
| One autonomy switch for everything A single on or off covering every task | One bad experience turns the whole feature off | One dial per task type | Autonomy dial |
| Silent irreversible action It already happened, with no preview and no undo | Trust rarely returns after this | Preview, undo, and an honest expiry | Intent preview |
| Notification flood Several agents each interrupting on their own schedule | People stop reading, including the one that mattered | One attention budget across every agent | Interruption budget |
| Chat as the container for everything Parallel and overnight work forced through one transcript | Nobody can say what is happening right now | A decision queue and a run board | Agent inbox |
| Anonymous fleets Agents with no name, owner or remit | Nobody can say who authorized what | Owner, scope, spend cap, review date | Agent registry |
| Confident guessing Ambiguity resolved quietly into something plausible | The error arrives finished and formatted | Three ways out: notify, ask, review | Escalation triad |