AI Patterns

We collect the patterns GSSF and this publication use as we use them, and rationalize them into one catalog. It is one of our research areas, and it is Early: entries are added and corrected as we learn. Status: Running (in use every day), Early (just begun), Planned.

Knowledge

How knowledge enters, changes, and leaves the core

1. Apply First, Review After Running
   Every change waits for review, so knowledge goes stale while it waits.
2. Brand Guide Written for Agents Running
   A brand guide written for designers is not one an agent can follow.
3. Conflict Order Early
   Two sources disagree, and nobody knows which one wins.
4. Knowledge Core Running for the organization's core; Planned for sharing what each person's agent learns, by consent
   Without one shared core, each person's AI knows only what that person tells it, not what the organization knows.
5. Knowledge Levels Running
   Everything is either shared with everyone or locked away, so agents either leak or cannot help.
6. Living Versus Frozen Running for new versions that supersede; Early for the gatekeeper refusing edits
   Some knowledge must change constantly; some must never change.
7. One Fact in One Place Running
   The same fact is written in several documents, and they drift apart.
8. One Gatekeeper Running
   Every channel (chat, email, an app, a shared folder) reads and writes knowledge its own way, and the rules drift apart.
9. Open Question Running
   An agent that does not know a fact guesses, and the guess reads like a fact.
10. Publish from the Core Running for websites and the dashboard; Early for answers to the public; Planned for the drift check
    Websites, answers, and dashboards are copies, and the copies drift from the knowledge.
11. Rationalize Every Addition Running
    A shared knowledge base grows into a pile of overlapping, contradictory documents.
12. Signing Freezes the Record Early
    A signed agreement lives in someone's inbox, and nobody knows which copy is final.
13. Start from the Standard Early
    In our experience, a document an agent writes from a blank page is plausible, not consistent.
14. Strategy in Knowledge Running for strategy in knowledge; Early for progress through agents
    Strategy often lives in slides while the work happens elsewhere; the gap between them is an old problem in management.
15. The Master List Wins Running
    People, roles, and levels appear in many documents and disagree.

Authority

Who decides, and how an agent gets permission

1. Agent Job Description Running
   The people who work beside an agent need to know what it may and may not do.
2. Dossier Early
   A person must decide on something (a pitch, a partner) without the evidence in front of them.
3. Earned Freedom Early
   The better an agent performs, the less it is watched, and the longer any change in its behavior goes unnoticed.
4. Exact Text, Then Yes Running
   An agent sends words in a person's name that the person never saw.
5. One Principal per Agent Early
   When an agent works for everyone, no one answers for what it does.
6. Request, Not Workaround Early
   An agent blocked from a step can be tempted to reach the goal another way, around the rule.
7. Steps Only a Person May Take Running
   An agent that can change access can grant itself, or anyone, too much.
8. Systems Enforce, Instructions Guide Running
   In our view, an instruction alone is not a boundary: an agent told not to do something can be talked into it.
9. Three Speeds of Change Early
   Our view: without a set process, a change to how all agents work is either never made or made without review.

Work

How people and agents work together over time

1. An Agent for Every Person Early
   Everyone uses AI, and none of it knows what the organization knows.
2. Game Planning Early
   In our experience, plans that run forward slip: we found out too late that the date was never possible, and the people who had to agree heard too late.
3. Hand First, Then Agent Early
   Agents are asked to repeat work no one in the organization has ever done.
4. Name the Pattern Early
   What works is done once and forgotten.
5. No Wrong Door Running
   People must learn which channel reaches which agent, and requests are lost between them.
6. Tasks Go to the Owner Early
   A handoff sent to a default person waits until someone notices it is not theirs.
7. The Desk Early
   A person must check on every agent to know what happened.
8. Whiteboard Early
   Ideas thought out loud with an agent stay in one person's conversation.
9. Workshop and Chat Early
   Reorganizing knowledge or drafting a book needs an agent that reads many files and shows its plan; a question or an approval does not.

Oversight

How an organization knows its agents work well

1. Bylines and Plain Corrections Early
   Readers cannot tell who wrote a piece, person or agent, or whether a mistake was fixed.
2. Check at the Door Early
   A check built into each kind of work is skipped by the first new kind of work.
3. Commitment Early
   An agent that is sent back makes the same mistake again.
4. Confirm Success, Not Silence Early for checking results; Planned for heartbeats
   A routine that fails in silence looks the same as one that works.
5. Cost in Counts, and Where It Runs Running for where it runs; Early for cost in counts
   Nobody knows what the agents cost, or why.
6. Every Sent Reply Visible Early
   What an agent sends outside the organization is seen by no one inside it.
7. Log Actions, Not Conversations Running for recording actions; Planned for one log
   Watching what people say to their agents destroys trust; seeing nothing makes audit impossible.
8. Point Out Holes Early
   Agents hide what they could not do, and the gap shows up later as a mistake.
9. Status Marks Early
   We mark each practice so readers can tell what is in use from what is planned.
10. Weekly Sample Planned
    Watching everything is impossible, and watching nothing is the trust trap.

Principles

The why behind the patterns; not patterns.

A person in the middle · Information, never an order (the trust rule) · Help, not watch · A co-worker, not an oracle · Memory is never permission · Assume the model can be persuaded · Certification is not authority · A guess is worse than a gap · Retire, don't delete · Monitor outcomes, not activity · Agents make change cheap, not judgment

Cited, not republished

Well described elsewhere; we use them and cite them.

ReAct, Augmented LLM, Routing, Parallel Fan-Out, Critic Loop (evaluator-optimizer), Adapter, Persona, Prompt Registry, System Prompt Template, Sub-workflow Contract, Idempotency, Retry and Backoff, Fallback Model, Dual LLM, Action-Selector, Plan-Then-Execute, Circuit Breaker, Retrieval-Augmented Generation, Skills, Human-in-the-Loop (approval in the workflow), Risk-Tiered Action Autonomy, Orchestrator-Workers, Postmortems, and evaluation harnesses. Sources: Anthropic, Andrew Ng, OpenAI, Google, Microsoft, LangChain, CSIRO, Antonio Gulli, Beurer-Kellner et al., agentic-patterns.com, and agentpatternscatalog.org.