AI development patterns
Real patterns in use at leading software organisations, drawn from first-hand engineering accounts, not vendor pitches or third-party journalism.
Source material: ai-dev
Environment and harness
-
Harness engineering
established Invest in the environment the agent works inside — feedback loops, docs, verification — rather than polishing each output by hand. The engineer's job shifts from writing code to designing environments, specifying intent, and building the loops that let agents check their own work. Models churn; the harness is the durable value.
- OpenAI — Harness engineering: leveraging Codex in an agent-first world
(Feb 2026)
notes - Datadog — Closing the Verification Loop
(Mar 2026)
notes - Shopify — Building an agentic harness that outlasts the model
(Jul 2026)
notes - Mitchell Hashimoto — My AI Adoption Journey
(Feb 2026)
notes - Brian Lovin (Notion) — Give your agent a laboratory
(Jan 2026)
notes
- OpenAI — Harness engineering: leveraging Codex in an agent-first world
(Feb 2026)
-
The agent-readable repository
established Encode decisions, conventions, and context as files in the repo that agents inherit automatically: AGENTS.md maps, structured docs directories, design decisions stored as agent-readable documentation and enforced by linters. Documentation stops being an afterthought and becomes part of the machinery.
- Block — AI-Assisted Development at Block
(Jan 2026)
notes - Vercel — Teaching agents product design at Vercel
(Jun 2026)
notes - OpenAI — Harness engineering
(Feb 2026)
notes - Anthropic — Building a C compiler with a team of parallel Claudes
(Feb 2026)
notes
- Block — AI-Assisted Development at Block
(Jan 2026)
-
Disposable sandboxed environments
established Agents get fresh, isolated, throwaway compute — cloud VMs and sandboxes spun up per task — instead of a developer's machine. Dangerous execution is contained, parallelism is free, and broken environments are discarded rather than repaired. By mid-2026 this looks like consensus infrastructure.
- Amp (Sourcegraph) — What I Want to Tell You About Orbs
(Aug 2026)
notes - Cursor — How we set up our cloud agent environment
(Jul 2026)
notes - Fly.io — The Design & Implementation of Sprites
(Jan 2026)
notes - Stripe — Minions — Part 2
(Feb 2026)
notes - Ramp — Why We Built Our Own Background Agent
(Jan 2026)
notes - David Crawshaw — Eight more months of agents
(Feb 2026)
notes
- Amp (Sourcegraph) — What I Want to Tell You About Orbs
(Aug 2026)
-
Agent-friendly is human-friendly
emerging Work done to make a codebase legible to agents — monorepo consolidation, standardised builds, readable docs, one-command environments — keeps turning out to help human engineers too. "If something is hard for me to read, it's hard for an agent to read."
- Shopify — Under the River
(May 2026)
notes - Vercel — Making Turborepo 96% faster
(Mar 2026)
notes - Stripe — Minions
(Feb 2026)
notes
- Shopify — Under the River
(May 2026)
Autonomy and review
-
The one-shot background agent
established Fire-and-forget agents: a task goes in — often from a Slack message — and a CI-passing pull request comes out, with no interaction in between. Humans review the result, not the process. At several companies these agents now author a large share of all merged PRs.
-
Specialist review panels
established Decompose code review into concern-specific sub-agents — security, performance, quality, docs — whose findings a coordinator dedupes and filters before anything reaches a human. The most advanced deployments grant full autonomous approval to a growing slice of PRs.
-
Verify, don't read
established When agents produce more code than humans can read, review capacity becomes the bottleneck. The response is to replace line-by-line reading with verification: near-perfect test suites, simulation testing, small enforced increments, and proof-of-functionality from longer validation runs.
-
Self-verification against real telemetry
emerging Close the agent's loop with production-grade evidence: run tests, read observability data, take screenshots, compare against recorded baselines. An agent that can check its own work against telemetry needs far less human babysitting.
- Ramp — Why We Built Our Own Background Agent
(Jan 2026)
notes - Brian Lovin (Notion) — Give your agent a laboratory
(Jan 2026)
notes - Vercel — Making Turborepo 96% faster
(Mar 2026)
notes
- Ramp — Why We Built Our Own Background Agent
(Jan 2026)
-
Security agents at scale
emerging Agentic pipelines that discover, validate, and patch vulnerabilities across whole codebases — with humans only at final review. The mirror image is defensive: AI review of inbound contributions, and containment of the agents themselves.
- Ramp — ~100 security issues fixed in 6 days with 0 humans
(Mar 2026)
notes - Shopify — An agentic harness that outlasts the model
(Jul 2026)
notes - Figma — How Figma Stays Ahead of Vulnerabilities With Agents
(Jul 2026)
notes - Block — CodeCrucible: LLM-driven SAST
(Jul 2026)
notes - Anthropic — How we contain Claude across products
(May 2026)
notes
- Ramp — ~100 security issues fixed in 6 days with 0 humans
(Mar 2026)
Orchestration and economics
-
Parallel fleets and swarms
emerging From one agent to many: fleets working a shared codebase with task locks, swarms implementing whole specifications, engineers running multiple changes concurrently. The coordination problems — locking, dedup, merge discipline — are the new engineering surface.
-
Model-tier economics
emerging Deliberately match model cost to task role: cheap models at high reasoning effort for the bulk of the work, frontier tokens reserved for coordination and judgment, and budgets enforced centrally to catch runaway loops. Token price turns out to be a poor predictor of task cost.
-
Evals on your own codebase
established Public benchmarks don't predict performance on your code. Teams build internal benchmarks from real tasks in their own repos and run them continuously — daily regression suites for agents, deterministic graders over LLM judges, and personal "back-pocket evals" at the individual scale.
-
Agents where the conversation is
established The agent's front door is the team chat, not the IDE. Slack-native agents triage incidents, open PRs from discussions, and work in public multiplayer sessions whose transcripts compound into shared knowledge.
- Shopify — Under the River
(May 2026)
notes - Duolingo — Triage, Debug, and Ship Code From Slack
(May 2026)
notes - Block — Buzz!
(Jul 2026)
notes - Stripe — Minions
(Feb 2026)
notes
- Shopify — Under the River
(May 2026)
Organisation and culture
-
Compounding knowledge layers
emerging Skills, corrections, and hard-won fixes are captured in version-controlled, code-reviewed repositories that every agent inherits — so the whole fleet gets better each time one agent learns something. Bad skills are the anti-pattern: they get the same things wrong every time.
- Shopify — Under the River
(May 2026)
notes - Fin (Intercom) — Claude Code: Good skills, bad skills
(Jul 2026)
notes - Replit — AI adoption starts with truth
(Aug 2026)
notes
- Shopify — Under the River
(May 2026)
-
Non-engineers shipping code
emerging With agents handling the mechanics, people outside engineering ship real changes: PMs implementing their own fixes, designers and support staff contributing via natural language, and half a company building internal tools.
-
The human as orchestrator-editor
established The senior practitioner's day shifts from writing code to directing, reading, and editing it — a 95:5 read-to-write ratio, delegation of what agents do well, and deliberate retention of the problems worth doing by hand.
- David Crawshaw — Eight more months of agents
(Feb 2026)
notes - Mitchell Hashimoto — My AI Adoption Journey
(Feb 2026)
notes - DHH — Promoting AI agents
(Jan 2026)
notes - Steve Yegge — Vibe Maintainer
(Mar 2026)
notes
- David Crawshaw — Eight more months of agents
(Feb 2026)
-
Known limits and failure modes
contested The honest counter-current: harness lock-in as models overfit to one dominant tool schema, over-trust in AI diagnoses, agent proactivity as a security liability, inflated adoption claims, and the case for keeping hand-written code where precise thinking matters.
- Armin Ronacher — Better Models: Worse Tools
(Jul 2026)
notes - Armin Ronacher — Building Pi With Pi
(May 2026)
notes - Simon Willison — Claude Fable is relentlessly proactive
(Jun 2026)
notes - Doug Turnbull — Why write code in 2026
(Jul 2026)
notes - Datadog — When an AI Agent Came Knocking
(Mar 2026)
notes
- Armin Ronacher — Better Models: Worse Tools
(Jul 2026)