Onomos @o is for autonomous Get in touch

Writing

Before the cliff: getting enterprise R&D ready for autonomous development

A ten-person company now ships tens of thousands of changes a month with a fleet of agents. Its founder says big companies will struggle to follow. Here is what it takes for an enterprise R&D organization to get there, one value stream at a time.

Onomos

None of it works unless you close the loop

On 29 October 2026 Jacob, the founder of Spawn, described how his ten-person company works. A permanent fleet of AI agents, each owning part of the company, shipped 77,000 substantial changes in 30 days. Pull requests are gone. A full test suite runs every 15 minutes. Changes are reviewed by agents that did not write them, and anything irreversible, such as a payment, passes through a small set of fixed gates. Of 123,548 changes over the past year, about 0.9% were reverted.

His own conclusion is the part enterprise leaders should read twice: big companies will struggle to adopt this, because none of it works unless you close the loop.

That is a statement about organizations, not models. The same models are available to every enterprise. What Spawn has, and most enterprise R&D does not, is a software development lifecycle built for agents. Agents read production directly. Every change is observable and reversible in seconds. Review is independent by design. Decisions people have already made are written down so agents can apply them.

The cliff comes one value stream at a time

David Shapiro calls this the automation cliff, and corrects a common picture of it. No company automates everything at once. Each value stream has its own cliff: nothing visible changes while the automation is built and tested, and then it takes the whole volume.

Because AI is not deterministic, each stream first has to earn its reliability, the “nines” between a demo that works most of the time and a process you can trust. Google DeepMind researchers made a related point this month: better individual models change the economy gradually, while better coordination between agents changes it abruptly.

For an enterprise, the question is therefore not whether to adopt agents. It is which value stream in R&D is ready next, and what it needs before it goes over.

What readiness means in practice

  1. Baseline the lifecycle. Measure lead time, deployment frequency, change failure rate and time to restore, so improvement is visible and so is regression.
  2. Sort the work. Place each task in one of four stages: Human-led, Agenticware (software helps, a person owns the outcome), delegated to an agent, or run by an agentic unit. Most tasks start on the left. That is fine.
  3. Close the loop for one value stream. Observability an agent can read, rollback in seconds, tests that run continuously, and a clear rule for what counts as done.
  4. Separate making from checking. No agent approves its own work. Review is done by a different agent or a person, and irreversible actions pass fixed gates.
  5. Write the decisions down. The answers your architects and product leads give over and over become principles agents can apply, with a route to a person when a question is new.
  6. Measure, then move the next stream.

None of this requires deploying an agent on day one, and all of it keeps its value if deployment waits: the same rules make human and mixed teams faster and clearer.

What we have learned

Onomos runs its own development this way on agent.ceo: agents in defined roles, completion only on evidence, no agent approving its own work, and escalation to a person when it matters. We have prepared enterprise engineering organizations for this transition, and the pattern repeats. The model is rarely the constraint. The lifecycle around it is.

Where to start

Pick one value stream in R&D, such as dependency upgrades, test maintenance, bug triage or a data pipeline. Baseline it, sort its tasks, and write the rules that would let an agent own it. That work is bounded, and it is the part Spawn’s founder says enterprises will find hardest.

If you want to see where your R&D stands, book a 30-minute call or write to aut@onomos.ai.

Sources

  1. Jacob (@jsnnsa) (2026). “How I Built the Most AI-Pilled Company in the World.” X, 29 October 2026. https://x.com/jsnnsa/status/2115623382721126826
  2. Shapiro, D. (2026). “The automation cliff.” YouTube. https://www.youtube.com/watch?v=W6b7MHwUBHk
  3. Imas, A., Koh, A. and Legg, S. (2026). “The Economics of Intelligence.” DeepMind Institute, 29 October 2026. The authors state their views are their own. https://institute.deepmind.com/essays/the-economics-of-intelligence/