提示：动态工作流：https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code 可用于让多个子代理并行分析大量数据，或对另一个代理的工作进行对抗性审查。在使用像 Claude Opus 或 Claude Fable 这样的模型时，可以说“分发多个子代理”或“使用工作流”。
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.
If you want to take a peek at the future of work, ask startups how they are operating today. So we did.
We spoke with more than a dozen fast-growing startups about how they use agentic coding tools to build products and scale their companies. These startups are changing the rules of who gets to build, what gets scrapped, and how to create a flywheel between how you build and what you build.
And they are shipping like organizations ten times their size.
In this guide, we'll dive into the unique deployments of these organizations to learn the rules they follow to ship fast and maintain their competitive advantage.
In doing so we'll also start to glean an answer to the question: what would it look like if an organization built their product development lifecycle with Claude Code from the ground up?
Tip: Only interested in the practical next steps? We've put a checklist at the end of this guide：#checklist that consolidates the key technical tips contained in each chapter.
Agentic coding lowers the barrier to entry, so the person who understands the problem can ship the first version of the fix.
Agentic coding lowers the barrier to entry for non-technical employees to build products. With Claude Code, you can create functional features without being fluent in a coding language or how to use an IDE.
For startup founders this has obvious advantages. For one, they don't have the headcount of their larger competitors so it's "all hands on deck." But it's not just raw capacity that founders are after–these non-technical members of the team bring domain expertise as well.
We heard the same thing from Dr. Thomas Kelly, co-founder and CEO of Heidi：#.
Saying "everyone ships" makes for a great LinkedIn post, but how does that work in reality? Is the marketing team approving pull requests? Is the legal team working through the intricacies of bisecting flaky tests?
The answer we got is that there is still a division of labor. Marketers still focus on marketing and developers still focus on developing. But the all important first step of getting an idea to working prototype, of going from 0 to 1, is open to everyone.
We also saw the most effective startups create mechanisms to make these contributions systemic rather than leaving it to chance or individual ambition.
It's one thing to create expectations for employees to use AI, it's another to give them access to Claude Code and the tools they need.
At Crosby, the team didn't bring lawyers to Claude Code, they brought Claude Code to the lawyers by connecting it to the tools and operating systems they were familiar with and worked in every day.
At some point, ideas need to be given the opportunity to be prioritized so that organizational resources can help bring them to market. That road is clear for product managers—it's their job after all—but not as clear for non-technical employees.
Clay creates quarterly reviews where prototypes are considered and can enter the formal roadmap. This is how a go-to-market team member at Clay built an autonomous agent that visits your websites, fills out your lead-capture forms, times how long it takes to respond, rates the experience, and generates a performance report.
Omni has a dedicated Slack channel for Claude generated prototypes with contributions from everyone including senior technical staff. They also practice the corollary of "everyone ships," which is "everyone talks with customers."
The line between "everyone ships" and "piecemeal" can be a thin one. Feature prototypes, whoever they come from, still need to be integrated into a product that feels like a cohesive whole. This is where skills, reusable instruction files that encode your team's standards and context, can help ensure development stays aligned even as the process becomes increasingly democratized.
"Anyone on the team can draft product components, marketing collateral or deck material from Claude Code using our design system as reference. AI that touches the product must clear a much higher bar, which Claude Code helps us meet with more precision," said Dr. Thomas Kelly, Heidi.
They can also get new developers and non-technical employees onboarded and up and running quickly.
Tip: Skills can be shared across the company using a directory：https://code.claude.com/docs/en/plugin-marketplaces so one employee's best practice can be instantly transferred to another. 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. For more information, read: Steering Claude Code: when to use CLAUDE.md, skills, hooks, and subagents：https://claude.com/blog/steering-claude-code-skills-hooks-rules-subagents-and-more.
Agents own the mechanical 80% of the lifecycle so engineers spend their time on the cases that actually need judgment.
All companies have sought to gain efficiencies through technology since the dawn of the industrial revolution, but these startups separated themselves by the speed and depth of their adoption.
These founders believe AI is an essential component of their mission. Many are explicit that agents own the mechanical 80% so engineers spend their time on the cases that actually need judgment.
Specifically, we saw AI more tightly integrated across their SDLC stages than others as well as more purpose built agents designed to take recurring tasks end-to-end. Let's look at a couple examples of both.
Many of these featured startups have implemented means of accelerating their teams' onboarding into their agentic coding processes. For example, at Emergent, Mukund told us, "on day one, a new hire bootstraps their entire dev setup by pointing Claude at the right markdown file. If Claude hits anything broken or out of date during onboarding, it updates that file."
Tip: Code Review：https://code.claude.com/docs/en/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).
These engineers need to be onboarded quickly because these teams ship fast.
At these organizations, Claude Code not only helps generate code, but reviews it too. "We run automated code reviews against our vetted technical and compliance frameworks, flagging critical issues and routing suggested changes to the right reviewers before anything ships," said Dr. Kelly of Heidi.
Some of these organizations have also built custom agents for code review, testing, and CI. These startups have placed considerable attention on building loops：https://claude.com/blog/getting-started-with-loops vs just deploying code.
"My favorite [agent] is the "Translucent code reviewer," which fans out across a change, reviews it from multiple angles, and synthesizes the results the way one of our senior engineers would but faster than any one person could," said Translucent founder Jack.
Clay "...built an agent that handles…bug triage, from first pass to suggesting code changes for fixes," said Kareem.
Another consistent pattern was that these startups were not only using agentic loops in Claude Code to accelerate their development efforts, but they were also creating agents to accelerate recurring and often tedious processes.
This was often routine work so that more attention could be focused on their competitive advantage, customer relationships, and on top-line growth. One of the most common processes we saw accelerated by Claude was self-service data analytics.
Nearly every one of these companies had some process in place so they could make quick decisions with fresh data, including unstructured data, that fuels the pivoting so essential in the life of a startup.
For example, Clay built an internal analytics agent and Heidi uses Claude Code to categorize customer and clinician feedback alongside usage data to surface signals that matter for product insights.
Both ClickHouse and Omni ship products that package this type of AI data analysis within them, all powered by Claude.
Other examples include summarizing thousands of legal documents with subagents (Crosby), sweeping claims data to flag anomalies across sites (Commure), and continuously mining hospital financial data for warning signs no analyst team could catch in time (Translucent).
Tip: Dynamic workflows：https://claude.com/blog/a-harness-for-every-task-dynamic-workflows-in-claude-code 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."
You can't automate a process unless you have a reliable means of monitoring and verifying the outcome.
This rule is the necessary corollary to Rule 2: Automate the tedium. You can't automate a process, unless you have a reliable means of monitoring and verifying the outcome.
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.
To be clear, none of these startups are having agents merge to main and hoping for the best. Many of them operate in highly regulated industries and require strong governance frameworks. Cainex is a particularly illustrative example of combining agents with deterministic checks to read medical records and generate codes that direct hospital billing.
"Here's the loop Claude Code runs for us. We process a batch with an agent, and our auditors review the output in an internal app. They don't just see the codes. They see the model's reasoning, and they comment on both….Everything is versioned and auditable," he said.
"Then Claude Code takes over. It reads the original predictions, along with every correction and comment, straight from the database. Each correction is tagged by the kind of code involved, so Claude Code knows whether it's looking at a diagnosis issue, a procedure issue, or another category, and it can go straight to the guidance that governs that specific kind of coding.
From there, it finds the part of the agent's instructions that produced the mistake and revises it, or writes new guidance when the case is genuinely new. Every change is made against a versioned set of instructions and tested against the records that failed. The rule we enforce: fix the principle, not the example," he continued.
"Then the back-test. A record can have more than one acceptable coding, so it's not a string match. The check combines semantic matching against our accepted sets with a judge that asks, 'Is this a real error or just a different valid path,' and Claude Code adds its own comparisons on top.
It runs the candidate change across a golden set plus random samples and surfaces any regressions before anything ships. What comes back is a short list: suggested edits, the records it couldn't resolve, and the questions it wants answered. Engineers spend their time on genuinely hard cases rather than the mechanical 80%," he said.
There are many generalized takeaways that founders can glean from this healthcare billing specific workflow.
For example, Cainex uses subject matter experts to routinely review and guide Claude's reasoning, and ensure that guidance becomes part of a self-improvement loop. However, those experts aren't there to fix example by example, their guidance is used as part of a self-improvement loop. As Uriah puts it "fix the principle, not the example."
The other takeaway is the diligence placed on maintaining a strong evaluation "golden set," or group of verified question answer pairs the team uses to verify the agent's accuracy. Every startup should maintain multiple sets of evals for their key use cases, and update them regularly, so they can prevent drift and evaluate future models.
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：https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents.
The final point Uriah makes is that this process can take some work. "It didn't start this clean. Our first version overfitted. It would 'fix' things by encoding the specific case, and we were accumulating patches instead of getting smarter. We changed the approach to force general principles and to cap how many specifics can enter a change at all."
Model capability keeps shifting underneath these teams, so very little is treated as permanent.
Many of these AI-native startups are in a state of constant reinvention.
AI is often at the heart of what they are building as well as how they are building it. Since model capability continuously evolves, groundbreaking features and critical scaffolding were discarded the minute they became sunk costs. Many of these organizations saw this constant rebuilding as part of their competitive advantage.
"What we do at Clay is you build it and then you build it again and then you build it again. And then the fourth time you build it, you know everything that's needed and you get it right. And so we don't necessarily throw away things. We just rebuild it: and this time with more clarity," said Kareem.
"A rebuild isn't done when the new path ships. It's done when the old path is gone. Teardown always lost the prioritization fight before: it's tedious and it ships no features," said Commure co-founder Tanay. "Now one of Commure's engineers just invokes a Claude skill to the tune of 'for every feature flag already released to everyone, open a PR removing it and the associated code,' then the engineer reviews what comes back. Migrations that used to eat a lot of dev cycles are now a plan and a fan out, done in a couple of hours."
扫码直达文章详情