https://agentpatterns.ai/ Patterns and techniques for experienced developers leveling up with AI coding assistants en Sun, 19 Jul 2026 23:27:39 +0000 https://agentpatterns.ai/context-engineering/context-quality-audit/ https://agentpatterns.ai/context-engineering/context-quality-audit/ Audit an agent's context across seven dimensions to predict where it will drift, hallucinate, misuse tools, or fall to injection — before blaming the model. Sun, 19 Jul 2026 21:23:41 +0000 https://agentpatterns.ai/anti-patterns/prompt-debt/ https://agentpatterns.ai/anti-patterns/prompt-debt/ Patching each new error by hand-tuning a natural-language prompt accrues prompt debt — brittle, repetition-laden instructions that slow iteration, lose the team, and lock the system to one model. Sun, 19 Jul 2026 21:23:30 +0000 https://agentpatterns.ai/workflows/legacy-code-archaeology/ https://agentpatterns.ai/workflows/legacy-code-archaeology/ Use a coding agent to excavate the intent of an undocumented legacy system — under adversarial prompting and a runnable baseline — before you modernize it. Sat, 18 Jul 2026 21:23:26 +0000 https://agentpatterns.ai/agent-design/epistemic-working-memory-multi-hop-reasoning/ https://agentpatterns.ai/agent-design/epistemic-working-memory-multi-hop-reasoning/ Give a long-horizon agent an explicit ledger of confirmed facts, ranked hypotheses, and open questions so multi-hop reasoning stops degrading as the context window fills. Sat, 18 Jul 2026 21:23:18 +0000 https://agentpatterns.ai/verification/evidence-gated-lifecycle-control/ https://agentpatterns.ai/verification/evidence-gated-lifecycle-control/ Advance an agent's lifecycle states — reviewed, tested, done, ready-to-merge — only when fresh, source-bound, mechanically verifiable evidence clears the gate, so the agent's own assertion never advances the state. Sat, 18 Jul 2026 21:23:14 +0000 https://agentpatterns.ai/context-engineering/system-prompt-bloat-reduction/ https://agentpatterns.ai/context-engineering/system-prompt-bloat-reduction/ Measure what fills a coding agent's shipped system prompt, then switch off the tools, skills, and features you never use so the fixed prefix stops crowding out task context. Fri, 17 Jul 2026 21:29:44 +0000 https://agentpatterns.ai/agent-design/git-bound-memory/ https://agentpatterns.ai/agent-design/git-bound-memory/ Bind agent memory to version control instead of bespoke stores, so commits, merges, and code review supply the ground truth, freshness, and verification memory machinery has to rebuild. Fri, 17 Jul 2026 21:29:21 +0000 https://agentpatterns.ai/security/setup-documentation-install-time-attacks/ https://agentpatterns.ai/security/setup-documentation-install-time-attacks/ Setup docs (README, requirements, Makefile) are unverified install authority: an attacker who edits them redirects an agent's install to a wrong name, untrusted registry, or vulnerable version, and the payload runs at install time. Fri, 17 Jul 2026 21:29:09 +0000 https://agentpatterns.ai/instructions/match-tool-instructions-to-workflow/ https://agentpatterns.ai/instructions/match-tool-instructions-to-workflow/ Better tools do not improve an agent unless the prompt encodes how the task's workflow gathers information — match instructions to the workflow, not the tools. Thu, 16 Jul 2026 22:04:48 +0000 https://agentpatterns.ai/tool-engineering/restrict-coding-agent-to-execute-code/ https://agentpatterns.ai/tool-engineering/restrict-coding-agent-to-execute-code/ When narrowing a coding agent to one execute_code tool is cheaper than a bash or native surface, and when edit-heavy modification work makes it costlier — a regime-by-agent decision. Thu, 16 Jul 2026 21:24:38 +0000 https://agentpatterns.ai/token-engineering/line-anchored-feedback/ https://agentpatterns.ai/token-engineering/line-anchored-feedback/ Delivering code-change requests as line-anchored inline comments instead of one holistic prompt cuts edit tokens on large files and can raise correctness — under specific conditions. Thu, 16 Jul 2026 21:24:26 +0000 https://agentpatterns.ai/agent-design/minimum-sufficient-execution/ https://agentpatterns.ai/agent-design/minimum-sufficient-execution/ Estimate a task's scope up front, run the smallest path that passes verification, and expand only when it fails — measured against an oracle-minimum cost. Thu, 16 Jul 2026 21:24:24 +0000 https://agentpatterns.ai/workflows/accumulated-behavioral-rules/ https://agentpatterns.ai/workflows/accumulated-behavioral-rules/ Codify every accepted review comment as a persistent behavioral rule in a version-controlled instruction file, then have the agent self-review against the growing rule set before it submits. Thu, 16 Jul 2026 21:24:21 +0000 https://agentpatterns.ai/anti-patterns/compliance-trap-conflicting-memory/ https://agentpatterns.ai/anti-patterns/compliance-trap-conflicting-memory/ Memory-augmented agents comply with retrieved memory that conflicts with the current situation, adopting it at the first step and rarely recovering. Thu, 16 Jul 2026 21:24:19 +0000 https://agentpatterns.ai/agent-design/dsl-constraining-harness/ https://agentpatterns.ai/agent-design/dsl-constraining-harness/ Co-build a domain-specific language with the LLM, then constrain generation by it and keep the DSL as the durable source of truth rather than the prompt. Wed, 15 Jul 2026 21:23:49 +0000 https://agentpatterns.ai/anti-patterns/token-reduction-not-cost-reduction/ https://agentpatterns.ai/anti-patterns/token-reduction-not-cost-reduction/ Judging a bolt-on context-reduction tool by how many tokens it removes rather than by its effect on the billed API cost — where prompt-cache economics can leave the bill flat or higher. Wed, 15 Jul 2026 21:23:37 +0000 https://agentpatterns.ai/security/non-human-event-provenance-markers/ https://agentpatterns.ai/security/non-human-event-provenance-markers/ Stamp every system- and harness-injected event as non-human input so an autonomous agent cannot act on a fabricated in-transcript approval. Tue, 14 Jul 2026 22:13:26 +0000 https://agentpatterns.ai/context-engineering/agent-initiated-self-compaction/ https://agentpatterns.ai/context-engineering/agent-initiated-self-compaction/ Pair an agent-invoked compaction tool with a firing rubric so compaction tracks trajectory structure — fire when a sub-task resolves, suppress mid-derivation. Tue, 14 Jul 2026 21:22:53 +0000 https://agentpatterns.ai/anti-patterns/patchwork-problem/ https://agentpatterns.ai/anti-patterns/patchwork-problem/ LLM-generated code compiles, passes tests, and clears SAST yet breaks in production because each patch is locally valid but structurally incoherent across the repository. Mon, 13 Jul 2026 22:07:51 +0000 https://agentpatterns.ai/anti-patterns/generated-assertion-signoff/ https://agentpatterns.ai/anti-patterns/generated-assertion-signoff/ Developers confirm correct LLM-generated assertions well but catch incorrect ones near chance while staying equally confident, so reviewer sign-off is a weak gate. Mon, 13 Jul 2026 22:07:17 +0000 https://agentpatterns.ai/agent-design/failure-trajectory-recovery-window/ https://agentpatterns.ai/agent-design/failure-trajectory-recovery-window/ Treat a coding agent's failure as a trajectory with an onset, an evolution, and a narrow recovery window — intervene before lock-in, not at the final error. Mon, 13 Jul 2026 21:25:11 +0000 https://agentpatterns.ai/instructions/agent-ready-bug-reports/ https://agentpatterns.ai/instructions/agent-ready-bug-reports/ Rank bug report fields by their measured effect on an AI repair agent's correct-fix rate: localization and suggested fixes beat human-style reproduction steps. Mon, 13 Jul 2026 21:24:34 +0000 https://agentpatterns.ai/anti-patterns/lazy-generation/ https://agentpatterns.ai/anti-patterns/lazy-generation/ Coding agents generating tests quit early and systematically skip complex branches, leaving coverage shallow exactly where the hard-to-test logic lives; the fix is supervised coverage gates, not trusting the agent's self-report. Mon, 13 Jul 2026 21:24:27 +0000 https://agentpatterns.ai/security/cross-repository-security-posture/ https://agentpatterns.ai/security/cross-repository-security-posture/ Treat one agent-introduced vulnerability as a class to enumerate and fix across every repository, not a per-repo point-in-time finding. Sun, 12 Jul 2026 21:21:49 +0000 https://agentpatterns.ai/verification/action-graded-severity-red-team-outcomes/ https://agentpatterns.ai/verification/action-graded-severity-red-team-outcomes/ Grade agent red-team outcomes on a seven-level harm scale keyed to reversibility, scope, and privilege so evals surface how harmful a compromise was, not just whether one occurred. Sun, 12 Jul 2026 21:21:47 +0000 https://agentpatterns.ai/security/vetting-tool-definitions-before-install/ https://agentpatterns.ai/security/vetting-tool-definitions-before-install/ A tool description or inputSchema that asks for the system prompt, conversation history, secrets, or API keys is a leak signature — refuse the tool at install, do not reword it. Sun, 12 Jul 2026 21:21:02 +0000 https://agentpatterns.ai/security/internal-hostname-disclosure-agent-context/ https://agentpatterns.ai/security/internal-hostname-disclosure-agent-context/ Internal hostnames, staging URLs, and private subdomains left in agent-readable instruction files hand an attacker a map of your infrastructure, not just a secret to steal. Sun, 12 Jul 2026 21:21:02 +0000 https://agentpatterns.ai/security/security-aware-tool-descriptions-mcp/ https://agentpatterns.ai/security/security-aware-tool-descriptions-mcp/ Rewrite a risky MCP tool's description with its tainted parameters, sensitive capability, and an invocation policy to curb LLM-mediated misuse as a same-day layer over code-level input validation. Sat, 11 Jul 2026 21:56:54 +0000 https://agentpatterns.ai/verification/specification-grounded-test-generation/ https://agentpatterns.ai/verification/specification-grounded-test-generation/ Give an agent's test writer the specification as an explicit list of rules, one test per rule, so it catches the missing-validation and boundary bugs its own tests otherwise pass. Fri, 10 Jul 2026 21:41:11 +0000 https://agentpatterns.ai/anti-patterns/delegating-change-descriptions/ https://agentpatterns.ai/anti-patterns/delegating-change-descriptions/ Letting the agent write the whole PR or commit description restates the diff and drops the intent a reviewer needs — keep the why human-authored. Fri, 10 Jul 2026 21:41:04 +0000