Work Open. Hold One Accountable. Give the Agents a Seat.
What if you told your entire development team to stop working, and to hand over their laptops to a dark software factory? What if it built, reviewed, tested, merged and deployed your product (100s of PRs a day) without human intervention, and it worked? I did that in September 2026. It was scary, beautiful (and expensive), and it left one question that no AI policy has answered: when the agents do the work, who is accountable, and how are decisions made?
My answer focuses on decision making in a mixed agents plus humans workforce: Work open. Hold one accountable. Intent is ours. Serve the decision. Know what we decided. Everybody has a seat. Record who signed; (WHISKER). One accountable human for every decision. A seat for every agent, on a mandate that can only narrow and never contains a secret. Every "yes" signed and recorded. Built on twenty years of Drupal issue-queue culture, tested at my own company, and frank about what is not solved.
What does "human in the loop" mean when the loop runs at machine speed and the human is asleep? It is the most repeated and least defined phrase in AI policy. Article 14 of the EU AI Act requires human oversight of high-risk systems; the human-factors literature going back to Bainbridge (1983) says that a human parked on top of an automated process to "oversee" it is exactly the arrangement that fails, because monitoring is the job people are worst at. This talk is about what the loop should actually look like once a dark software factory can do the middle of software work, drawn from running a company that way since August 2026.
The claim: software engineering has become U-shaped. Agents specify, implement, test and merge. Humans decide what to build and why at one end, and accept or refuse what was built at the other. Sovereignty lives at the two ends.
I will present a way of working, WHISKER, that my company runs on today:
- Work in the open; decide where anyone can read along.
- Hold one human accountable for every decision.
- Intent is ours; implementation is the machine's.
- Serve the decision: every step before and after is in its service.
- Know what we already decided; agents inherit it.
- Everybody, humans and agents, has a seat.
- Record what was decided, what they saw, who signed.
For this track I will spend the most time on the four lines that are about control:
- Hold one accountable. Every decision has exactly one accountable human, derived from the org chart rather than asserted by whoever opened it, with a live veto over the machine. Why a signer without a real stop is what Madeleine Elish calls a moral crumple zone, and how we avoid building one.
- Everybody has a seat. A channel treated as a responsibility chart with seats for humans and agents. An agent's seat is a mandate: a signed credential listing what it may do, where, and how far it may delegate. Authority can only narrow down the chain, revocation cascades, and the agent never holds the underlying secret, so a prompt-injected agent has nothing to leak. Shown as a working proof, built on the same capability ideas behind UCAN and current IETF work on workload identity, and framed honestly: the standards are individual drafts, the semantics are what we adopt now.
- Record who signed. What a decision record must contain to be defended later: the question, the options, the default that would have applied, what the decider actually read, who held each seat, the signature. Why silence is never approval, and the cheap countermeasure we adopted after an agent fabricated a "ratified" approval on a real architecture decision.
- Model sovereignty at the merge boundary. Nothing trusts the agent's word. Tests run again, the change is compared to the plan it claims, and it is read by a reviewer on a different model family from the one that wrote it. That rule makes the way of working provider-neutral by construction and gives open and locally hosted models a structural seat as the independent reviewer, rather than a nice-to-have.
I will also be plain about the surveillance risk in "work in the open," and how we handle it: decisions and evidence are durable and public; presence and status are ephemeral and treated as private data; agents may carry what is happening elsewhere but may never act on it without confirmation.
The Drupal community solved an early version of all this before anyone said human-in-the-loop: decisions in a public queue, a review state that only a non-author could set, the why recorded where anyone could read it. The talk carries that culture into a world where most participants in the queue are not human, and names the queue's pathologies as the missing pieces: a review that sits for a year is silence pretending to be a state, and the committer bottleneck is Bainbridge's monitor.
Spec Kitty is the open source tool we built to do this to ourselves and it appears as the case study, but every mechanism is stated in terms a Drupal team could implement with an issue queue, a commit trailer, a CI job and a text file, and every research claim in it has been checked against its source.
Learning outcomes
Attendees will leave able to:
- Give "human in the loop" an operational definition: where the human sits, what they must see, and what they sign.
- Bound an agent's authority so that it can only narrow, can be revoked in one step, and never holds a secret.
- Run a channel or issue thread as an explicit responsibility chart that includes agents, and answer the four seat questions for any decision.
- State what a decision record must contain to be defended to a regulator, an auditor, or a customer.
- Explain why a second model family at the merge boundary is a sovereignty control, not a quality nicety.
Audience and level
Skilled to expert. For maintainers, tech leads, product owners, CTOs and policy people who have watched an agent do a week of work in an afternoon and wondered who is accountable for it. No prior experience with AI agents required; familiarity with the Drupal issue queue helps; no cryptography assumed.
Robert Douglass is the founder and CEO of Spec Kitty, where a fleet of AI agents builds, reviews and merges the company's software while the humans decide and accept. He has been a Drupal contributor since the mid-2000s: author of the Apache Solr and Memcache modules, co-author of one of the first Drupal books, member of the Drupal Association's General Assembly, and Drupal's representative in Google Summer of Code. He was an early employee at Acquia, later a vice president at Platform.sh, and an early investor in Fission, where the UCAN capability model was born. He lives in Bonn.