Lutflow in Claude Startups: how we work with Claude in engineering
Lutflow was accepted into Claude Startups, Anthropic's program for early-stage founders building on Claude, on October 6, 2026. Here is what the program is according to Anthropic, what we received and, above all, how we work with Claude in engineering: separate roles, claims checked against evidence, and final decisions in people's hands.
The news
On October 6, 2026, Anthropic accepted Lutflow into Claude Startups, its program for early-stage founders building on Claude.
What Claude Startups is, according to AnthropicCITED
Anthropic presents Claude Startups as its program for early-stage founders building on Claude (program page). In the announcement of its expansion, dated October 6, 2026, it lists what accepted startups get:
- a one-time USD 1,000 Claude API credit;
- a free year of Claude Team with up to five Premium seats, for organizations new to Team;
- the Claude Startup Stack: discounts and credits on products built on Claude, offered by independent third-party companies, not by Anthropic;
- office hours with Anthropic’s Applied AI team, events and community.
According to the program page, the credits apply to the first-party Claude API through Claude Console; they cannot be used for Claude on AWS, Google Cloud or Azure.
What Lutflow received
According to the program panel in Lutflow’s account:
- 12 months of Claude Team, with up to USD 625 per month off (two-seat minimum).
- USD 1,000 in Claude API credits, which expire 180 days after they are granted.
- Access to the Applied AI sessions: live group sessions with Anthropic’s Applied AI team, every other week, each covering a topic and including an AMA.
- Access to the Claude Startup Stack: exclusive offers for members from products built on Claude. They are offered by independent third-party companies, not by Anthropic.
How we work with Claude in engineering
Lutflow is a small team building a measuring instrument. If the instrument overstates, it stops being useful. That is why the way we use AI agents in engineering is designed to avoid mistakes more than to go fast. We describe it as a practice, with generic examples.
Separate roles
We do not ask one agent for everything. We work with distinct roles, each in its own session and with its own context:
- Executor: makes the change on a branch, runs the tests and opens a draft pull request. It does not merge or publish.
- Independent reviewer: reviews the change adversarially, without the executor’s context. It looks for edge cases, results that pass for the wrong reason and claims the code does not support, and writes down its verdict.
- Supervisor: coordinates, checks every delivery against the evidence and prepares the decisions people make.
Every claim is checked against an artifact
“The tests pass” is accepted with the run’s log, not with the sentence. “The number is X” is accepted if it can be traced back to the measurement that produced it, with its label. Before a cost number reaches a slide or this blog, it is checked against the measurement report, and a person decides explicitly what gets published and with which label. If something cannot be verified, it is written as “unavailable” or it is not published.
Least privilege
Agents do not handle credentials: they neither create nor copy them. Keys are issued and held by people. When a step needs access, for example reading a cluster’s state, it uses an identity with just the permissions for that task and a short lifetime.
A dedicated publishing channel and human merges
Agents publish through a single path: their own branches and draft pull requests. Merging, publishing a package or cutting a release requires a person's explicit approval. This article arrived that way: as a draft a person reviews before publishing it.
Independent verification before publishing or acting in the cloud
Before something is published or cloud infrastructure is touched, a second pair of eyes checks it: another agent in the reviewer role, or a person. Cloud actions also require a person’s explicit approval, a cost and time cap, and a teardown confirmed by re-reading the state, not by assuming it.
What this is not
- Claude Startups membership describes program participation. Lutflow remains responsible for its product and its published numbers.
- Agents do not replace decisions: people decide what gets published, what gets merged and what is done in the cloud.
Claude and Anthropic are trademarks of Anthropic.
Frequently Asked Questions
What are separate roles for?+
So that nobody reviews their own work. The reviewer does not share the executor's context, so it does not inherit its assumptions, and the supervisor accepts a delivery only with evidence.
Who is responsible for what Lutflow publishes?+
Lutflow is responsible for what it publishes. Membership describes participation in the program.