Free · MIT · v2.0 · For people who run AI agents in production

Action Building is the operating practice for ultracoding: one person directing many AI agents to ship production software.

Agents now write code faster than any person can review it. What decides whether the product holds up is how the work is run: contracts fixed before agents start, checks proven to fail, one home for every fact, and features that ship live on the default path. Action Building packages those rules as skills your agents read. Proven in enterprise delivery: tenant data isolation, encrypted patient data, an independent review and a rollback plan on every release, all run by agents in production. If you know what the product needs, this is how you direct the agents that build it.

Get the pack → See the skills

Ultracoding

Ultracoding is one person directing many agents to release software on demand

In May 2026 Anthropic added ultracode to Claude Code: one word in a prompt, and Claude writes a workflow that runs dozens of agents in parallel. Bun's creator used it to port about 750,000 lines of code from Zig to Rust in eleven days, with 99.8% of the existing tests passing. Builders have started calling this way of working ultracoding.

Running agents in parallel is now a setting. Deciding what is worth building, keeping dozens of agents consistent with one another, and proving that a release is safe for customers are still the operator's job, and that is where most teams stall. Others are working on the same problem. Paul-Antoine Tual describes ultracoding as four practices: orchestrate, challenge, test and decide. Paul Kuo argues that the human work left is judgment about what's worth the compute, and taste about when a result is good enough to count as done. Action Building is our version, run in production.

An ultracoder owns the whole job for one part of the product. They understand what the customer needs, plan the release, direct the agents that build it, and decide when it ships. A fresh-context agent reviews every merge, which is the work a second engineer used to do.

Agile grouped work into two-week sprints because coordinating people was slow: planning meetings, handoffs between product, engineering and QA, and a queue for code review. When agents build and one person holds those roles, the sprint has nothing left to batch. We call this post-Agile. Work ships as micro-releases on demand: each is a small, versioned change with its own review, tests, release notes and rollback plan, released whenever it is ready. Our core team averaged about nine releases a day in the third quarter of 2026.

A two-week sprint
1 release
The same two weeks, ultracoding
a micro-release whenever one is ready
Illustrative: one tick per release at the Q3 2026 average of about nine releases a day. Each tick carries its own review, tests and rollback plan.

The rest of this page covers the rules that make that pace safe.

Further reading

Problems that start when agents outnumber your team

With these rules in place, the speed of the agents reaches your customers. For product owners, turning customer needs into working software becomes a system the team can trust: requests move from conversation to production faster, and every release arrives ready for customers to use.

How a small team or builder ships faster and with higher quality than a large one

Five scalable practices your agents follow from the first day

A large engineering team spreads auth, schema, deploys, migrations, reviews and security across many people and outdated processes. Action Building writes that knowledge into skills your agents read, so one person or a small team streamlines all of it.

01

Buy the plumbing

Use managed services for everything already solved.

Auth, database, hosting, payments and email come from managed services, for example Clerk, Supabase, Vercel, Stripe and Resend. Your agents integrate them and never rebuild them.

02

Install the rules on Day 0

Put the rules in place before the first feature.

That means the schema manifest, database access policies, version discipline, and a review before every release. Rules added after the first incident cost a rebuild.

03

Stay conversational

Run every operation from the conversation or the app.

Connectors replace the terminal. Maintenance jobs run from an admin screen with a dry run first. No secret passes through chat. The conversation history is the audit trail.

04

Fix the design and plan, then run agents in parallel

Lock the design and the release plan before any agent starts.

One orchestrator and many build agents work from the same plan and shared test fixtures, with a fresh review at every merge.

05

Turn every lesson into an improvement in quality

When something surprises you, make it a repeatable test.

Lessons become rules, rules become tests that flag you when your system is broken, and tests become skills the next agent reads.

In the downloads The Ultracoding Pack holds 74 skills, 74 pattern pages and 11 templates, plus the ultracode workflow and the team guide. The Security pack adds 7 skills and 23 automated checks.

The proof

Releases rose almost fourfold in one quarter, and the release controls held

These numbers come from the release history of a multi-tenant healthcare platform that handles patient data, built by a small core team directing AI agents under the rules in this pack. Every release carried a version number, customer-facing release notes, a deploy checklist with smoke tests and rollback triggers, and a fresh-context code review.

Speed, Q3 2026

822

releases

3.7x the prior quarter (223)

About 9

releases a day

The quarter's average

2,531

changes shipped

4.5x the prior quarter (558)

Quality, held at that pace

About 11,000

unit test cases

Each proven to fail before it is trusted, run on every build

43

rules

Each tied to a recorded incident

763

maintenance runs

All from inside the app, dry run first, none from a command line

Figures are for Q3 2026 (July to September). The 2024 DORA report defines its elite tier as deploying on demand, a tier fewer than one in five organizations reach.

RYS

Who made this

Written by a repeat founder who runs his companies' software with AI agents

Robin Y Smith is a chemist-turned-serial-founder with nearly 30 years of building software and life-science companies. He invented the electronic lab notebook and holds 23 issued US patents. He founded and sold Synthematix (acquired by Symyx, 2005) and ArtusLabs (acquired by PerkinElmer, 2011), the two companies that pioneered the digital lab notebook category. He later founded Orig3n/Seaport Diagnostics (acquired by KKR, 2021). EY Entrepreneur of the Year New England finalist, 2018. Today he runs Wonderlab Bio and Darby from Boston, where a small core team builds and operates their software by directing AI agents under the rules in this pack.

"In every market I'm in, we compete against companies with much bigger engineering teams, and the team that gets customers what they asked for first usually wins. Agents let our small team move at that speed. Our post-Agile method of ultracoding micro-releases on demand gives us a huge competitive advantage."
LinkedIn X / @robinysmith Wonderlab Bio Darby

How it works

Install the pack, then run every release through the same loop

Getting started

  1. Install the plugin in Claude, or add the skill folder to any AI tool that reads files.
  2. Tell your agent what the product does, who uses it and what it runs on, for a new project or one already running.
  3. The agent sets up the rules before the first feature: schema manifest, database access policies, version and deploy notes, and the review step.
  4. You approve the plan, and the agent starts building.

The release loop

  1. Plan. An architecture and engineering plan, checked against the schema manifest.
  2. Draw and audit. Object diagrams and user flows, each checked for simplicity, plus mockups where the screens change.
  3. Pressure test. Simulated tests or a small harness prove the approach before the build.
  4. Build in parallel. An orchestrator gives each agent a fixed contract and its own lane.
  5. Merge and review. Every merge gets a fresh-context review and the full test suite.
  6. Ship. A versioned release through your normal deploy path (a pull request or a zip from chat), with a deploy checklist and a rollback plan, live on the default path.
  7. Learn. Whatever surprised you becomes a rule and a test the next agent reads.
The release loop Seven steps in a circle: plan, draw and audit, pressure test, build in parallel, merge and review, ship, learn. The learn step feeds the next plan. 1 Plan 2 Draw andaudit 3 Pressuretest 4 Build inparallel 5 Merge andreview 6 Ship 7 Learn Learn feeds the next plan

A one-line fix and a month-long program run through the same loop. Only the size of the plan changes.

For teams

How to train your team to work with agents and these skills

On an agent-built product, the roles that used to need separate people collapse into one. A single ultracoder holds the customer need, directs the agents and owns the release, while a fresh-context agent does the review a second engineer used to do. A team is several ultracoders working in parallel, each owning a part of the product, all working from the same rules.

Once the rules are shared, a new ultracoder's agents work to the same standard as yours from their first session.

The team guide ships inside the pack.

Why we call it Action Building

In 1952 Harold Rosenberg called the canvas an arena in which to act

Rosenberg coined the term Action Painting for Jackson Pollock and the painters around him. In his account the finished painting was the record of an event, the trace of what happened while the painter worked, rather than a picture planned in advance and filled in. For years critics said anyone could throw paint. Later analysis of the drip paintings found fractal structure, a result physicists still argue about, and it made people look again.

Building with agents works the same way. The product gets made inside the conversations with the agents, and the codebase is the record of what they did. When many agents work at once, that record is worth keeping only if every move followed the same rules.

Get the pack

Download the Ultracoding Pack and the security checks, free under the MIT license

Leave your email and we'll send the download links. If you tick the box, we'll also write when a new version ships, no more than twice a month. We don't share your address.

For Claude

One-click install. Claude replaces an older version in place.

Action Building 2.0 Ultracoding Pack
74 skills, 74 pattern pages and 11 templates, the ultracode workflow and the team guide. Plugin, v2.0.0, 564 KB.
Download plugin
Security pack
7 skills and 23 automated checks, including prompt injection screening. Plugin, v1.2.0, 187 KB.
Download plugin

For other AI tools

Plain markdown skill folders. Put them in your repository and commit, so every session reads the same rules.

Action Building 2.0 Ultracoding Pack
Skill folder, v2.0.0, 590 KB.
Download zip
Security pack
Skill folder and CI checks, v1.2.0, 212 KB. Runs in GitHub Actions.
Download zip

The Ultracoding Pack. The skills, pattern pages, templates and rules your agents read, plus the team guide. Install it as a plugin in Claude, or add the skill folder to any AI tool that reads files.

The Security pack. Automated checks your pipeline runs on every change: secrets in code, database access policies on every table, background work that can be lost, webhook signatures, injected HTML in emails, and prompt injection screening. Runs in GitHub Actions. Caught real critical findings on its first field run.

Install or update the packs, step by step

Version 2.0 updates the Cold Start Pack in place under its new name. Every earlier skill is still in it, some of them revised. A first install and an update follow the same steps.

In the Claude app

  1. If you use Safari, first open Safari, then Settings, then General, and untick Open "safe" files after downloading. Otherwise Safari unzips the plugin into a folder, and Claude needs the zip.
  2. Open the download link in your email. Under "For Claude", click Download the plugin on the Ultracoding Pack card. The file action-building-ultracoding-pack-plugin-v2.0.0.zip lands in your Downloads folder. Leave it zipped.
  3. In Claude, click your name at the bottom of the left sidebar, then Settings.
  4. In the left menu, under Customize, click Plugins.
  5. Click + Add at the top right, then Upload plugin.
  6. On the "Upload a plugin" page, choose the zip from step 2. The preview shows action-building-cold-start. Click Upload.
  7. If you have an earlier version installed, Claude asks "Replace existing plugin?". Click Replace.
  8. Under "Created by you", click Action Building 2.0 Ultracoding Pack. Its page shows version 2.0.0.
  9. Do steps 2 and 5 to 8 again for the Security pack, using action-building-security-plugin-v1.2.0.zip.
  10. Click + New to start a new session. A session that was already open keeps the old version.

In other AI tools

  1. Open the download link in your email. Under "For other AI tools", click Download the zip on the Ultracoding Pack card. Your browser may unzip it into a folder for you; either is fine.
  2. Unzip it if needed. You get one folder named for the pack and version.
  3. In your repository, delete the old pack folder and put the new folder in its place.
  4. Commit the change, so every agent session reads the same version.
  5. Start a new session in your AI tool.

Check: in the new session, ask "Which Action Building version am I on?" The answer should be 2.0.

Who the Ultracoding Pack is for

Starting from zero

If you're building your first product with agents, the pack starts you on the same rules a production team uses. Its start skill walks you through Day 0: it asks what you're building, recommends managed services so you don't rebuild solved problems, sets up your database design and access rules, and wires a deploy pipeline with a rollback. You begin with the habits that usually arrive only after the first incident, so your prototype can grow into a production system without a rebuild.

FAQ

  • Is it free? Yes, under the MIT license. Use it, change it and share it with credit.
  • Why do you need my email? To send the download links. Updates come only if you tick the box.
  • Does it work outside Claude? Yes. The skills are plain markdown files, so any AI tool that reads files can use them. The one-click plugin install works in Claude.
  • Is it only for healthcare software? No. The rules were proven on a healthcare platform, where mistakes are expensive, but they apply to any production system with real users and real data.
  • Do I need to be an engineer? No. Product owners and first-time builders start with the Day-0 walkthrough. Engineers get the orchestration, verification and release rules.
  • How does it handle security? The separate Security pack runs automated checks in your build pipeline on every change, including screening for prompt injection.
  • Is a book or paid version coming? A book and a hosted analyzer are planned. The pack stays free.
  • Where do I get help? Email hello@action-building.com, connect with Robin on LinkedIn, or message him on X. Contributions are welcome.

Release as often as your customers need, with a small team

Our core team releases several times a day against competitors many times our size. Install the pack, and your agents plan, build in parallel, review and ship by the same rules.

Download the Ultracoding Pack →