This guide is also available for download — the same five rules, founder insights, and checklist, laid out for reading offline or sharing with your team.
Tip: Only interested in the practical next steps? We've put a checklist at the end of this guide that consolidates the key technical tips contained in each chapter.
01
Everyone ships
Agentic coding lowers the barrier to entry, so the person who understands the problem can ship the first version of the fix.
エージェントコーディングは、非技術系従業員が製品を構築する参入障壁を低くします。 Claude Code を使用すれば、コーディング言語や IDE の使い方に精通していなくても機能的な内容を作成できます。
"Not only were engineers shipping much more, but non-technical people (like me) were also suddenly shipping UI changes and other product improvements."Mads Lunau Liechti · co-founder, Parahelp
"Claude Code changed what it meant to be a lawyer at Crosby. The lawyers have the best product insights, because they are the users. It's been amazing to watch them cook."Ryan Daniels · co-founder and CEO, Crosby
Heidi の共同創業者兼 CEO であるトーマス・ケリー博士からも同じことを聞きました。
"For us, Claude Code solved the broken telephone problem. The way a new idea used to move through a team was the person with the idea tells a PM, who tells a designer, who then tells an engineer… and inevitably the essence of the idea gets lost in that chain. By the time something shipped, it often didn't resemble what the person had in mind. And it took weeks. Claude Code collapses that chain. The person who actually understands the problem can ship a PR bringing in designers and engineers for the parts where their expertise matters."Dr. Thomas Kelly · co-founder and CEO, Heidi
従業員が AI を活用するよう期待を高めることと、Claude Code や必要なツールにアクセスできるようにすることは別です。
"We're actually not running away from [having non-technical employees contribute], we're going towards it. Our take is every role is becoming an engineering role because you can build software for it… so we hire people who are tinkerers, who are interested in building"Kareem Amin · co-founder and CEO, Clay
Crosby では弁護士を Claude Code に導入するのではなく、弁護士が慣れ親しんで日常的に使用しているツールやオペレーティングシステムに接続して、弁護士に提供しました。
Tip: Claude can't understand what it can't see. One of the most effective ways to extend Claude's value is to connect it to sources of truth and the tools your team uses every day.
MCP is an open source standard for AI-tool integrations that give Claude Code access to your tools, databases, and APIs. Explore adding these connections whenever your team finds itself copying and pasting information from a tool into Claude.
Connecting via CLI can be more token-efficient when a mature command-line tool already exists (gh, kubectl, bq, psql) and you want Claude working against the same ground truth your engineers do.
Omni には Claude が生成したプロトタイプ向けの専用の Slack チャネルがあり、シニアテクニカルスタッフを含む全員が貢献しています。 また、「全員が成果を成果を上げる」という必然的な原則、つまり「全員がお客様と対話する」も実践しています。
Even though engineers don't naturally gravitate toward customer calls, Omni deliberately puts them in front of customers because it closes the feedback loop faster.Chris Merrick · co-founder and CTO, Omni
「チームの誰もが当社のデザインシステムを参考にして、Claude Code から製品コンポーネント、マーケティング資料、資料を作成できます。 製品に携わる AI ははるかに高い基準をクリアする必要があり、Claude Code はより正確な基準を満たすのに役立ちます」と、Heidi のトーマス・ケリー博士は述べています。
また、新しい開発者や非技術系従業員をオンボーディングし、迅速に稼働させることもできます。
"...we also have a GitHub repo of Claude Code skills which works as a shared knowledge base to quickly bootstrap a Claude Code session with known Emergent details like database [and data warehouse] location, some schema [information], overall company context….instead of trying to be perfect here, it is ok to live with slightly outdated context files as long as the agent can quickly verify and course correct."Mukund Jha · co-founder and CEO, Emergent
"Our engineers use Claude Code to spin up an in-house marketplace of specialized internal agents, organized by role, so engineering, delivery, and sales each get tools built for how they actually work."Jack O'Hara · founder and CEO, Translucent
"Everyone's racing to build AI products. Far fewer are rebuilding how their company actually runs. The second one is the bigger unlock. Artemis Security runs as an AI-native company, not a company that happens to use AI. This supercharges our velocity and allows us to help customers stop attacks at machine speed."Shachar Hirshberg · co-founder and CEO, Artemis Security
具体的には、SDLC のステージで他よりも AI がより緊密に統合されており、繰り返し発生するタスクをエンドツーエンドで処理するように設計されたエージェントがより多くなっています。 両方の例をいくつか見てみましょう。
AI ネイティブの SDLC
これらの特集のスタートアップの多くは、チームのエージェントコーディングプロセスへのオンボーディングを迅速化する手段を実装しています。 たとえば Emergent では、Mukund が次のように述べています。「初日に新規採用者が適切な markdown ファイルを Claude に示すことで、開発セットアップ全体を自律的に管理します。 オンボーディング中に Claude が壊れたり古い内容を検出した場合は、そのファイルを更新します。」
Tip:Code Review (research preview) is a managed multi-agent service in Claude Code. It runs an automated review pass on PRs in the repos you enable. You can manually fix the finding and push, or close the loop by commenting @Claude on the finding (if you've set up and configured GitHub Actions).
"Engineers here are orchestrating agent fleets, shipping fixes to production data problems the same day they're found, and running multiple PRs in flight simultaneously. One engineer ran a ~13-ticket initiative with Claude subagents in parallel, each owning a ticket and its PR."Tanay Tandon · CEO and founder, Commure
Tip: For the last several months Claude Tag has been the on-call first responder for CI/CD failures at Anthropic. Claude authored the first situation report in every recent incident that had one, typically publishing its first analysis within 15 minutes.
Claude Tag has its own service account and access to the tools an Anthropic CI engineer needs such as Datadog or Grafana. Standing instructions are in markdown files as skills, committed in a GitHub repository. This way multiple teammates can iterate on them and we can manage changes just like we do code.
Claude タグが Slack のオンコールスレッドを理解し、チャネルで進捗状況を報告します。
This was most pronounced at ClickHouse, where co-founder and CTO Alexey Milovidov reported the database company had turned nearly every SDLC stage into an autonomous loop. Two purpose-built agents designed to fix flaky tests and find missing test coverage are now the #2 and #3 contributors to the ClickHouse repo. A separate family of agents handles operations, and the team uses Claude Code to build and iterate on those agents themselves.
エージェントでプロセスを迅速化
もう 1 つの一貫したパターンとして、これらのスタートアップが開発努力を加速するために Claude Code のエージェントループを活用するだけでなく、繰り返し発生するしばしば面倒なプロセスを加速するためにエージェントを構築していたことです。
これは多くの場合で日常的な作業であり、競争優位性、顧客関係、収益の成長により集中できるようにしました。 Claude によって加速化された最も一般的なプロセスの 1 つは、セルフサービスのデータ分析でした。
Tip:Dynamic workflows can be used to fan multiple subagents to analyze large amounts of data in parallel or to conduct an adversarial review of another agent's work. When using a model like Claude Opus or Claude Fable say "fan out multiple subagents," or "use a workflow."
03
Trust, but verify
You can't automate a process unless you have a reliable means of monitoring and verifying the outcome.
Artemis Security co-founder Dan Shiebler said their increased deployment speed only works…"because we've invested deeply in testing infrastructure, codebase organization, and team knowledge systems that let agents ship end to end. This is the flywheel we've built with Claude: structure your codebase, knowledge base, and team the right way, and every contribution compounds."Dan Shiebler · co-founder, Artemis Security
"Early on we gave Claude full autonomy and it did what AI does. It shipped plausible code fast. The problem was it drifted from our architecture in ways that looked right but weren't. So we…wrote down every invariant. How we frame problems. What has to be true no matter what. How to prove something works instead of trusting a confident answer. 567 lines of how this team thinks."Victor Hunt · co-founder and CEO, Zingage
Tip: Put what can't change in CLAUDE.md at the root of your repo. Claude reads it at the start of every session, so your architecture rules, security boundaries, and non-negotiables travel with every session.
明確にしておくと、これらのスタートアップのいずれもエージェントを main にマージさせており、最善の成果を期待していません。 その多くは規制の厳しい業界で事業を展開しており、強力なガバナンスのフレームワークを必要とします。 Cainex はエージェントと決定的なチェックを組み合わせ、医療記録を読み取り、病院の請求書を導くコードを生成するという点で特に好例です。
"In medical coding, a wrong code isn't a typo. It's a billing and compliance event. That one fact governs how we build."Uriah Israel · co-founder and CTO, Cainex
Tip: Loops are agents that repeat cycles of work until a stop condition is met. They can be effective ways to use Claude Code for more autonomous or long-horizon work. You can use skills to define what criteria the agent needs to meet (the more clearly defined the better) and have the agent iterate until it reaches its goal.
For example, many organizations create flaky test agents, or loops, because the stop condition is clear and self-contained: the agent can verify its own fix by rerunning the test until it passes.
"[Claude Code has] also transformed how we manage model velocity. New video and image models arrive constantly. Each requires new skills, evaluations, routing logic, and production testing before deployment. Claude Code has compressed that cycle from days to hours, allowing us to identify issues in production and deploy fixes in the same session….When you're competing against companies with 10x the headcount, that kind of leverage changes everything."Alex Mashrabov · co-founder and CEO, Higgsfield
Tip: When teams first start building agents, they can get surprisingly far through a combination of manual testing, dogfooding, and intuition. The breaking point often comes when users report the agent feels worse after changes, and the team is "flying blind" with no way to verify except to guess and check. Teams can't distinguish real regressions from noise, automatically test changes against hundreds of scenarios before shipping, or measure improvements. For more information read: Demystifying evals for AI agents.
Tip: AI agents are not deterministic, but a lot of highly regulated work requires processes to be done the same way every time. Claude Code has features that can help combine frontier intelligence with deterministic processes.
Hooks are user-defined commands that fire at fixed points in Claude Code's lifecycle and can serve as hard gates. They execute every time regardless of what the model decides. For example they can be used to block a write that fails a lint, require a test pass before commit, or strip secrets before anything leaves the sandbox.
Dynamic workflows orchestrate subagents with deterministic sequencing, separate context windows, and focused goals. /goal is helpful for long complex tasks where Claude may prematurely call the job done, prefer its own findings when reviewing, and drift from its original goals.
04
Build for rebuilding
Model capability keeps shifting underneath these teams, so very little is treated as permanent.
Tip: Use git worktrees to run a rebuild in an isolated copy of the repo while the current version stays untouched. Claude Code can spin one up for you — you get v2 running next to v1, run your evals against both, and only merge when the new one wins. This is what makes "build it four times" cheap.
2026 年 5 月の「Code with Claude 」イベントで、Harvey の応用 AI 責任者 Niko Grupen が、新しいモデル機能(緊急推論、エージェントの自動化、プランニング、オーケストレーション)の各々に、プラットフォームの完全なアーキテクチャの再構築が必要であることについて講演しました。
"If you asked me six months ago what our architecture looks like, I'd give a fundamentally different answer from what it looks like today. If we hadn't been willing to say 'Hey, we need to scrap this and go agent native' we simply could not have these capabilities in our platform right now."Niko Grupen · Head of Applied AI, Harvey
Tip: For non-trivial rewrites, start Claude Code in plan mode (--plan or hit Shift+Tab). Claude will explore the codebase and propose the rebuild approach before writing any code — you approve or redirect. It's the cheapest place to catch a rebuild that's about to drift from your architecture.
05
Prototype, dogfood, productionize
Building with AI helps these startups create disruptive products with AI — the flywheel at the heart of their process.
これらのスタートアップの多くは、開発プロセスの中心に重要なフライホイールを設置しています。 AI で構築することは、AI で破壊的な製品を生み出すのに役立ちます。
「当社は SQL コンソールのエージェントや AI SRE など、チームが直接対話できる独自の AI エージェントを[自社の製品に]構築しました。 当社では Claude Code を使用して、これらのエージェント自体を構築し、反復的に運用しています。 お客様の AI 体験を支えるツールは、一部が AI で構築されています」と、ClickHouse の Alexey 氏は述べています。
チェックリスト
このガイドでは多くの分野を網羅しました。 以下は重要なヒントを 1 つのページにまとめたものです。
Chapter 1: Everyone ships
Claude can't understand what it can't see. Connect it to sources of truth and the tools your team uses every day via MCP or CLI.Create a company plugin marketplace so one employee's best practice can be instantly transferred to another via a skill. Use CLAUDE.md files in each subdirectory of your repo for coding conventions specific to that subdirectory that apply every time. Use skills for on-demand procedural workflows.
Chapter 2: Automate Tedium
Set up Code Review (research preview) on a repo for an automated review pass on PRs.Make Claude Tag (public beta) part of your CI/CD on-call response and bug triage.Dynamic workflows can be used to fan multiple subagents to analyze large amounts of data in parallel or to conduct an adversarial review of another agent's work.
Chapter 3: Trust, but verify
Put what can't change in CLAUDE.md at the root of your repo.Use loops, agents that repeat cycles of work until a stop condition is met, for more autonomous or long-horizon work.Establish a process for creating and maintaining agent evaluations.Hooks are user-defined commands that fire at fixed points in Claude Code's lifecycle and can serve as hard gates. Use these when components of the work need to be deterministic.
Chapter 4: Build for rebuilding
Use git worktrees to run a rebuild in an isolated copy of the repo while the current version stays untouched. This is what makes "build it four times" cheap.For non-trivial rewrites, start Claude Code in plan mode (/plan or hit Shift+Tab). Claude will explore the codebase and propose the rebuild approach before writing any code — you approve or redirect. It's the cheapest place to catch a rebuild that's about to drift from your architecture.
フロンティアに立つスタートアップはフロンティアで構築する
これらの洞察は最先端の分野で構築を進めている同僚から得られたものです。実用的かつ実用的だと感じていただけたら幸いです。 Claude スタートアップコミュニティは常に、インスピレーション、ベストプラクティス、アドバイスの源です。 このコミュニティには、以下の方法で参加できます。