Agent Usage Patterns

English

日本語

English

日本語

Appearance

Agent Usage PatternsA catalogue of patterns for how people use AI agents

Explaining the same premises over again every time. The model says it "fixed" something, but it isn't fixed. A large deliverable lands all at once, too much to review. If any of this sounds familiar, each symptom already has a named pattern for addressing it. Every claim carries an evidence label, so readers can judge its weight for themselves.

Find by Symptom
Usage Guide

Context Management

Instruction Layering and Scoping · Rules Ranked by Binding Force · Session Boundary Design · Keeping the Prefix Stable · Resident Context Inventory · Matching Thinking Depth to the Task

Delegation & Parallelization

Read-and-Discard Isolation · The Break-Even Point of Delegation

Verification & Trust

Separating Mechanical Gates from Semantic Gates · Computation as Code · Failure as Institutional Memory · Strong Gates for Unattended Paths · Bounding Permissions by Structure · Making the Verification Target Explicit

Human Checkpoints

Preemptive Plan Review · One Authoritative Source for Instructions · Diff Review and Incremental Adoption

Patterns Included ​

Listed by category.

Context Management ​

PatternSummaryLast Updated
Instruction Layering and ScopingSeparates instructions that always apply from ones scoped to just this moment. It works not because the total volume shrinks, but because only what's relevant remains2026-07-22
Rules Ranked by Binding ForceWhen every rule carries the same weight, important constraints get buried under mere preferences2026-07-22
Session Boundary DesignWhere to discard, and what to carry forward2026-07-22
Keeping the Prefix StablePut what doesn't change up front, and what does change after it2026-07-22
Resident Context InventoryCount what's paying rent on every turn2026-07-22
Matching Thinking Depth to the TaskThinking isn't free2026-07-22

Delegation & Parallelization ​

PatternSummaryLast Updated
Read-and-Discard IsolationOffload work that reads a lot but returns little to a separate context. What gets discarded is the process of reading; the result comes back as a summary2026-07-22
The Break-Even Point of DelegationDelegation carries a price tag measured in multiples2026-07-22

Verification & Trust ​

PatternSummaryLast Updated
Separating Mechanical Gates from Semantic GatesDon't leave to the model what a machine can already judge2026-07-22
Computation as CodeRoute reproducible computation to code. What disappears is error, not effort — answers become deterministic and checkable2026-07-22
Failure as Institutional MemoryRecord past incidents concretely, not as abstract admonitions. Writing down what the model already does just adds noise2026-07-22
Strong Gates for Unattended PathsThe less a human is watching a path, the more it should be bound by structure2026-07-22
Bounding Permissions by StructureInstructions get forgotten; structure doesn't2026-07-22
Making the Verification Target ExplicitHand over what to verify as a contract2026-07-22

Human Checkpoints ​

PatternSummaryLast Updated
Preemptive Plan ReviewA plan before execution is the cheapest deliverable to fix2026-07-22
One Authoritative Source for InstructionsWith no priority order defined, don't create conflicts in the first place2026-07-22
Diff Review and Incremental AdoptionHow the work is received is itself where a human intervenes2026-07-22

Scope and Approach ​

  • Scope — Patterns for how humans use AI agents. Each entry pairs a tool-agnostic core with a GitHub Copilot in Practice section.
  • Audience — The top priority is letting outside readers find a pattern from their own symptom.
  • Evidence — Every claim carries an evidence label. The goal isn't to prove rigor — it's to let readers judge the weight of a claim for themselves.
LabelMeaning
OfficialA vendor's primary source. Includes URL and retrieval date
ResearchPublished academic or third-party research
MeasuredA result the author observed in their own environment. Not written as a general law
UndocumentedA behavior observed that the official docs don't describe. Written as a delta against the official description
Absence ConfirmedChecked a primary source and confirmed the description is absent. States only that the description is absent, not that the behavior itself is denied
FolkloreCirculates in the community but no primary source can be traced. Noted as untraceable when included

This catalogue is community-driven, not first-party official documentation. It doesn't compete with official docs on rigor. Its reason for existing is that it can write what official docs structurally cannot — undocumented behavior, results checked firsthand, contradictions between docs, and the fact that something was searched for and not found.

The last one needs explaining. Official docs can't declare their own absence. Only someone searching from outside can write "there's no description of this feature anywhere." The fact that a check was made and came up empty isn't discarded here — it's kept as [Absence Confirmed].

For Readers Looking for Design Patterns ​

Technical architectures for building agents — ReAct, CodeAct, Agentic RAG, Self-Reflection, Multi-Agent Planner — are out of scope for this catalogue. The names are close enough to cause confusion, so it's worth stating plainly here. Readers looking for those will find the Agent Design Pattern Catalogue (academic, 18 patterns, retrieved 2026-07-16) a better fit. What this catalogue fills is the gap that catalogue leaves — usage patterns, the ways people use agents.