The Pattern Library
AI agent design patterns
An AI agent design pattern is a recurring interaction between a person and software that acts. The person delegates a task, approves an action, interrupts a run, reviews the result, stops the agent, or picks the work up afterwards; the agent shows, asks, waits, reports, or carries on. Each product with an agent in it is reinventing these interactions on its own. This library gives each one a name, a definition, an anatomy of the parts a screen needs, the tensions teams hit, and a component you can install.
The library is built from dated, sourced field notes. A pattern enters as a candidate when two notes from different products show the same interaction and moves to emerging when a third does; each morning the pipeline attaches the day's notes to the patterns they evidence and records the drift, and a status changes only when a person agrees. Every pattern page lists its notes with dates and sources, so you can check the claim rather than take it. Read where each of the 25 patterns came from →
Momentum
SIGHTINGS / TRAILING 90 DAYS
Published, annotated sightings attached to each pattern, trailing 90 days (Jul 17 – Oct 15). None of these patterns had a sighting in the prior 90 days (Apr 18 – Jul 17). n = 196.
View as table
| Pattern | Trailing 90d | Prior 90d |
|---|---|---|
| Observability | 40 | 0 |
| Approval | 36 | 0 |
| Permissions | 34 | 0 |
| Context | 19 | 0 |
| Review | 18 | 0 |
| Confidence Signaling | 17 | 0 |
| Delegation | 16 | 0 |
| Recovery | 16 | 0 |
SIGHTINGS / WEEK · SHARED 0–13 SCALE
Weekly counts of published, annotated sightings per pattern over the last 12 weeks (Jul 30 – Oct 15), every row on one scale; the dot is the latest week. Hover a week to read every pattern at once. n = 196.
View as table
| Pattern | Jul 30 | Aug 6 | Aug 13 | Aug 20 | Aug 27 | Sep 3 | Sep 10 | Sep 17 | Sep 24 | Oct 1 | Oct 8 | Oct 15 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Observability | 0 | 0 | 0 | 0 | 1 | 8 | 6 | 13 | 2 | 2 | 5 | 3 |
| Approval | 0 | 0 | 0 | 0 | 0 | 9 | 4 | 5 | 2 | 3 | 7 | 6 |
| Permissions | 0 | 0 | 1 | 0 | 1 | 9 | 4 | 6 | 2 | 4 | 3 | 4 |
| Context | 0 | 0 | 0 | 0 | 0 | 5 | 2 | 3 | 3 | 1 | 2 | 2 |
| Review | 0 | 0 | 0 | 0 | 0 | 2 | 2 | 5 | 0 | 3 | 2 | 4 |
| Confidence Signaling | 0 | 0 | 0 | 0 | 0 | 3 | 2 | 0 | 1 | 5 | 2 | 4 |
| Delegation | 0 | 0 | 1 | 0 | 1 | 4 | 2 | 4 | 1 | 0 | 2 | 0 |
| Recovery | 0 | 0 | 0 | 0 | 1 | 2 | 1 | 2 | 2 | 3 | 2 | 3 |
Coverage & Divergence
Which products show which interactions, computed from anatomy-level sightings. Anthropic first, OpenAI as the standing counterpart. Full map →
| Claude Code | MCP + MCP Apps | Apps SDK (apps in ChatGPT) | ChatGPT | ChatGPT Atlas | ChatGPT agent mode | Codex | |
|---|---|---|---|---|---|---|---|
| Approval | |||||||
| Delegation | |||||||
| Planning | |||||||
| Progress | |||||||
| Human Handoff | |||||||
| Interruption | |||||||
| Recovery | |||||||
| Agent Identity | |||||||
| Memory | |||||||
| Generative UI |
⬤ ring = one pattern's anatomy · 1 segment = 1 anatomy part · filled segment = part evidenced for that surface · hover a cell for its parts and last evidence · blank = no sighting yet
How to read a pattern
- The problem line. One sentence on what the person needs, product-agnostic. If it isn't your problem, it isn't your pattern.
- The maturity status. Watching → emerging → established → contested → fading, re-scored from sighting velocity. Methodology.
- Field notes. Every claim traces to a dated source; the side-by-side shows how each vendor handles each part of the interaction.
- Get the code. Patterns with components install via the shadcn registry. Install & MCP.
Watching
Candidate interactions accumulating sightings. Three or more from two independent vendors and they get a page.
- Confidence Signaling 17 sightings Knowing how much to trust a particular answer.
- Reach Preview 17 sightings Sizing the blast radius of an agent before it starts, and the ground it has covered since.
- Emergency Stop 16 sightings Stopping agents fast, in each place they run, and knowing they stopped.
- Off-Brief Alert 15 sightings Telling a person fast when an agent does something the brief did not cover.
- Budget Scoping 13 sightings Setting limits on spend, time, or reach before handing work over.
- Comparison View 10 sightings Answering "better or worse than what, and for whom" after a prompt change.
- Eval Set Authoring 9 sightings Turning "I know it when I see it" into examples and a standard a test can use.
- Release Decision 9 sightings Deciding whether a prompt change ships, with the evidence in view.
- Case Drilldown 8 sightings Judging one answer, and deciding whether the judge got it right.
- Production Drift 8 sightings Knowing when quality slips in production, measured the same way as the tests.
- Multi-Agent Roster 6 sightings Keeping a team of agents legible: who exists, who is doing what.
- Eval Run Setup 4 sightings Choosing what to test and knowing what it will cost before pressing run.
- Ambient Status 3 sightings Knowing what background agents are doing without opening them.