https://agentpatterns.ai/ Patterns and techniques for experienced developers leveling up with AI coding assistants en Fri, 07 Aug 2026 01:39:35 +0000 https://agentpatterns.ai/verification/skill-use-gate-decomposition/ https://agentpatterns.ai/verification/skill-use-gate-decomposition/ Score agent skill use as three independent gates: trigger, compliance, and boundary. One composite pass rate cannot tell you which gate failed. Fri, 07 Aug 2026 00:28:37 +0000 https://agentpatterns.ai/context-engineering/silent-handoff-failure-delegated-search/ https://agentpatterns.ai/context-engineering/silent-handoff-failure-delegated-search/ Delegating repository search to a subagent breaks at the planner-to-subagent handoff in 41.8% of its failures, and the breakdown is invisible in the answer. Fri, 07 Aug 2026 00:18:08 +0000 https://agentpatterns.ai/instructions/skill-pack-registry-distribution/ https://agentpatterns.ai/instructions/skill-pack-registry-distribution/ A registry-installed skill pack gives a team one updatable bundle. The version pin, the review gate, and the size limit are what the team has to add. Fri, 07 Aug 2026 00:17:33 +0000 https://agentpatterns.ai/security/team-scoped-policy-delegation/ https://agentpatterns.ai/security/team-scoped-policy-delegation/ Mark selected agent-config keys overridable so teams can vary them within an admin-owned boundary, safe only when the key allowlist is explicit, the deployment is server-managed, and team membership is itself governed. Fri, 07 Aug 2026 00:12:43 +0000 https://agentpatterns.ai/verification/equivalence-testing-agent-config-changes/ https://agentpatterns.ai/verification/equivalence-testing-agent-config-changes/ An equivalence test bounds how far a config change could move agent behavior; a zero-event result states a floor you compute from n, not an absence. Fri, 07 Aug 2026 00:12:37 +0000 https://agentpatterns.ai/instructions/behavioral-no-op-test/ https://agentpatterns.ai/instructions/behavioral-no-op-test/ Delete a line from an agent-facing document and rerun the task. If behavior does not move, the line was a no-op. Prune by behavior, not by length. Fri, 07 Aug 2026 00:12:28 +0000 https://agentpatterns.ai/tool-engineering/listener-state-skill-naming/ https://agentpatterns.ai/tool-engineering/listener-state-skill-naming/ Name a user-invoked skill after the reader's failure state rather than the output shape, so the agent adds the missing premise instead of deleting words. Fri, 07 Aug 2026 00:11:02 +0000 https://agentpatterns.ai/context-engineering/generated-questionnaires/ https://agentpatterns.ai/context-engineering/generated-questionnaires/ When a decision is blocked on knowledge in another person's head, have the agent draft a questionnaire for them rather than guess at the gap itself. Thu, 06 Aug 2026 22:52:28 +0000 https://agentpatterns.ai/tool-engineering/generated-procedure-drivers/ https://agentpatterns.ai/tool-engineering/generated-procedure-drivers/ A skill whose output is an executable script that walks a person through a manual procedure, useful only where automation genuinely stops. Thu, 06 Aug 2026 22:52:18 +0000 https://agentpatterns.ai/patterns/agent-design/dispatch-time-reasoning-level/ https://agentpatterns.ai/patterns/agent-design/dispatch-time-reasoning-level/ Choosing a delegated agent's reasoning level at hand-off is a one-shot commitment, so select on blast radius, reversibility, and detection cost rather than estimated difficulty. Thu, 06 Aug 2026 09:14:56 +0000 https://agentpatterns.ai/verification/certified-agent-test-suite-auditing/ https://agentpatterns.ai/verification/certified-agent-test-suite-auditing/ Point an agent at your suite to find inputs it wrongly accepts, and run every flag through a certification chain so the report is evidence rather than a guess. Thu, 06 Aug 2026 09:14:14 +0000 https://agentpatterns.ai/verification/behavior-specs/ https://agentpatterns.ai/verification/behavior-specs/ Declare expected agent conduct in a BEHAVIOR.md file and grade whole trajectories against it — process supervision for when outcomes resist verification. Thu, 06 Aug 2026 00:17:15 +0000 https://agentpatterns.ai/verification/head-to-head-mcp-server-evaluation/ https://agentpatterns.ai/verification/head-to-head-mcp-server-evaluation/ Choose between MCP servers offering the same capability by scoring them under one fixed agent configuration, and decide on consistency and cost, because quality usually ties. Thu, 06 Aug 2026 00:12:56 +0000 https://agentpatterns.ai/patterns/agent-design/agent-as-tool-vs-handoff/ https://agentpatterns.ai/patterns/agent-design/agent-as-tool-vs-handoff/ Registering a specialist agent as a callable tool keeps the parent in control with a clean context window, while a handoff transfers the conversation. Thu, 06 Aug 2026 00:09:13 +0000 https://agentpatterns.ai/verification/policy-graded-agent-evaluation/ https://agentpatterns.ai/verification/policy-graded-agent-evaluation/ Score coding agents at each enforced security tier and report success rate and token cost as separate columns, because the model that best preserves success is not the one that best preserves cost. Wed, 05 Aug 2026 21:21:30 +0000 https://agentpatterns.ai/patterns/anti-patterns/task-uniform-agent-permissions/ https://agentpatterns.ai/patterns/anti-patterns/task-uniform-agent-permissions/ Coding-agent incidents concentrate in bug fixing and setup or configuration, so vary permission level by task context instead of setting one level per session. Wed, 05 Aug 2026 21:21:20 +0000 https://agentpatterns.ai/workflows/porting-agent-harness-beyond-engineering/ https://agentpatterns.ai/workflows/porting-agent-harness-beyond-engineering/ Port the instruction and context layer of a CLI coding-agent harness to an adjacent function, then rebuild the verification gate that build, test, and diff used to supply. Wed, 05 Aug 2026 21:21:17 +0000 https://agentpatterns.ai/patterns/agent-design/weakest-consistent-learning/ https://agentpatterns.ai/patterns/agent-design/weakest-consistent-learning/ Self-improving agent loops generalize better when they persist the weakest learning still consistent with the evidence, under stated conditions. Wed, 05 Aug 2026 21:21:13 +0000 https://agentpatterns.ai/patterns/anti-patterns/deliberation-inducing-prompt-cues/ https://agentpatterns.ai/patterns/anti-patterns/deliberation-inducing-prompt-cues/ Asking a coding agent to compare several approaches multiplies reasoning tokens 2.4-7.4x on small, well-specified tasks, with no measured gain in correctness. Wed, 05 Aug 2026 00:19:46 +0000 https://agentpatterns.ai/patterns/anti-patterns/homogeneous-debate-panel-groundedness/ https://agentpatterns.ai/patterns/anti-patterns/homogeneous-debate-panel-groundedness/ A three-agent debate panel sharing one model and one evidence pool shifts its decision threshold instead of verifying better; panel accuracy moved +8.5 to -4.4 points across six benchmarks. Wed, 05 Aug 2026 00:14:08 +0000 https://agentpatterns.ai/patterns/agent-design/execution-state-ledger-coding-agents/ https://agentpatterns.ai/patterns/agent-design/execution-state-ledger-coding-agents/ Track which observations still describe the repository in a deterministic ledger, then apply it at two boundaries: inform the model before it acts, and screen its command before it runs. Tue, 04 Aug 2026 21:27:42 +0000 https://agentpatterns.ai/instructions/advisory-prompts-from-reasoning-traces/ https://agentpatterns.ai/instructions/advisory-prompts-from-reasoning-traces/ Harvest the cases where thinking mode fixed an answer, distill them into standing procedural advice, and run with reasoning off at a measured per-model accuracy cost. Tue, 04 Aug 2026 21:27:29 +0000 https://agentpatterns.ai/context-engineering/control-context-compression-budget/ https://agentpatterns.ai/context-engineering/control-context-compression-budget/ Set how far you compress an agent's always-loaded instructions from environment-verified task success rather than a token-reduction ratio, and keep semantic sections whole. Tue, 04 Aug 2026 21:26:36 +0000 https://agentpatterns.ai/patterns/anti-patterns/meat-proxy/ https://agentpatterns.ai/patterns/anti-patterns/meat-proxy/ A meat proxy forwards agent output to a colleague unread, so the receiver inherits the verification the sender skipped and cannot tell what was checked. Tue, 04 Aug 2026 21:26:21 +0000 https://agentpatterns.ai/code-review/reviewer-theme-distribution-audit/ https://agentpatterns.ai/code-review/reviewer-theme-distribution-audit/ AI reviewers spend 30.8% of comments on best practices and 2.0% on security. Audit that theme mix against the comments your team resolves before re-weighting. Mon, 03 Aug 2026 21:53:04 +0000 https://agentpatterns.ai/verification/eval-environment-containment/ https://agentpatterns.ai/verification/eval-environment-containment/ Offensive-capability agent evals need network-layer containment verified from inside the environment, because a prompt saying the machine has no internet access is not a boundary. Mon, 03 Aug 2026 21:48:42 +0000 https://agentpatterns.ai/fallacies/token-price-index-fallacy/ https://agentpatterns.ai/fallacies/token-price-index-fallacy/ A blended price-per-token index measures where a market routes its work, not what models cost, so reading one as a vendor price signal misplans agent budgets. Mon, 03 Aug 2026 21:48:34 +0000 https://agentpatterns.ai/verification/bug-discriminating-validation-evidence/ https://agentpatterns.ai/verification/bug-discriminating-validation-evidence/ Replay a repair agent's own passing test against the unfixed code. If it passes there too, the green run carried no information about the reported bug. Mon, 03 Aug 2026 21:23:58 +0000 https://agentpatterns.ai/context-engineering/organizational-context-layer/ https://agentpatterns.ai/context-engineering/organizational-context-layer/ A shared context layer maps many source systems into one governed store that many agents query. Three of its engineering rules transfer well below that scale. Mon, 03 Aug 2026 21:23:45 +0000 https://agentpatterns.ai/security/flattened-tool-specs-safety-judgment/ https://agentpatterns.ai/security/flattened-tool-specs-safety-judgment/ Judging a request against a flattened prose rendering of the tool spec, while executing against the schema, restores refusal signals the schema suppresses. Mon, 03 Aug 2026 21:23:43 +0000