The attention budget your AI agent orchestration plan is missing

By the Cerevisor team

Tomorrow, August 14, Claude Code flips its default. New sessions on Pro, Max, and Team plans start in auto mode, which means the agent proceeds without asking a person to approve each action unless the action looks irreversible, destructive, or aimed outside its workspace. The Register covered the change on August 10, and the same three days brought a toolkit that packages one investor’s personal setup into 23 specialized agents for solo founders, a cloud service that keeps agent jobs running after the laptop lid closes, and an AWS runtime that lets agent workloads run continuously for up to 14 days, up from an 8-hour limit.

Four separate announcements, one direction: more agents, running longer, interrupting the person less. The output math in every pitch is compelling. The attention math is missing entirely.

The week AI orchestration tools stopped asking for permission

Start with why the permission prompts are going away, because the reasoning is honest and a little uncomfortable. Anthropic ran a controlled study with 1,053 paid testers and found that the human checkpoint was mostly ceremony.

Those are the vendor’s own numbers about its own feature, so read them with that in mind. But the supporting figure rings true from every deployment I have watched:

So the checkpoint that interrupted people constantly was catching almost nothing, and the machine checks catch more. Fine. I believe it. The mistake would be concluding that the person is now free. What actually happened is that the person’s job changed shape: from approving individual actions to supervising whole agents, plural. AI agent orchestration, running several agents as one coordinated system where a plan decides which agent does what and in what order, is exactly the discipline of managing that plural. And the number of agents per person is climbing fast. The gstack toolkit that shipped August 10, covered in that day’s LLM Daily, bundles 23 named agent roles, from Engineering Manager to Release Manager to QA, and pitches solo founders on shipping “like a team of twenty.” Claude Code itself now defaults to 20 concurrent subagents per session, per a Digital Applied breakdown from late July. The always-on launches in the opening came from the AI Agent Store’s news log for the same week.

A team of twenty, one set of eyes. That sentence should appear somewhere in the pitch deck.

What building an AI agent orchestration platform taught us about parallel agents

Cerevisor is a visual AI agent orchestration platform, so we run into the plural-agents problem as a matter of daily engineering, and I want to share three specifics from our own executor, because they say something about where the whole category’s protections actually point.

First, the shape of a run. Cerevisor executes workflows in waves: a wave is a group of agents that run at the same time, and the next wave waits until the current one finishes. Inside a wave, we cap how many agents talk to the model provider at once at four, dispatching a wide wave in chunks of four rather than all at once. Here is the honest part: that cap exists because an unbounded fan-out blows through provider rate limits and the resulting failed calls waste wall-clock time retrying. It protects the API relationship. It was never designed to protect the human watching the run.

Second, a failure mode we had to design out. When several agents share one workspace in the same wave, the executor cannot always tell who wrote a file. A write detected by diffing the workspace could get credited to an agent whose parallel sibling actually touched the path, which would raise a conflict alarm blaming the wrong agent. The fix was to only credit an ambiguous write to an agent when the file is one of that agent’s own declared outputs. We also rate-limit the real alarms: when two agents genuinely write the same file in the same wave, the conflict alert fires once per file per wave, so a noisy writer cannot flood the log.

Sit with that for a second. A deterministic system, with full visibility into every event, needed special engineering just to attribute actions correctly across four parallel agents and to keep its own alerts from overwhelming the reader. That is the easy version of the problem. The hard version runs on wetware, has no event log, and is expected to do it for twenty.

Attention residue, the tax no AI agent orchestration framework can waive

The research on what parallel supervision costs a person is older than any AI agent orchestration framework, and it is not encouraging about the fleet model. The anchor study is Sophie Leroy’s 2009 paper in Organizational Behavior and Human Decision Processes, a peer-reviewed set of lab experiments, so I am citing seventeen-year-old work and saying so plainly. Leroy measured what she named attention residue: when a person switches away from a task, part of their mind keeps working on it, and the residue is strongest when the task was left unfinished. Her participants performed measurably worse on the next task when the previous one was incomplete.

Now map that onto delegation. Every agent you hand work to creates an open loop: a task that is officially in progress and unfinished by design, sometimes for hours, this week for up to 14 days. The pitch says “describe it and walk away.” Leroy’s data says the walking away is not free, because the unfinished thing follows the person into whatever they do next. Twenty parallel agents is twenty open loops held by one head.

The permission prompt was never the real cost of supervision. The open loop was, and auto mode does not close a single one.

There is newer evidence that supervision is substantial work rather than a residual. A June 2026 interview study on arXiv, from Dhanorkar, Passi, and Vorvoreanu, spoke with 17 experienced developers using software agents. It is a small qualitative preprint, not a measurement study, but its core finding is structural: oversight is not just reviewing output at the end. The authors identify four distinct kinds of oversight work, including setting up controls before the run, co-planning, and monitoring in real time. People are doing a whole job here. It just does not appear on any org chart or in any per-seat price.

Open loops per operator: the number to set before you approve a fifth agent

Here is what I would actually do with this, because the situation is very manageable once it is named.

Pick a number for concurrent agent streams per operator, out loud, the way we picked four for provider calls. The literature will not hand you the exact value, but it points at small: working-memory research puts actively attended items around three to four, and Leroy’s work says the cost compounds with unfinished tasks. Whether your number is three or six matters less than the fact that it is chosen, visible, and revisited, instead of drifting upward one “just one more agent” at a time.

Then design for loop closure, not just for throughput. An open loop stops taxing attention when it genuinely closes, which means checkpoints where a run ends in a decision, not a glance. This is most of why we built Cerevisor’s run view the way we did: waves end, results land in one place, and the operator processes a batch and closes it, rather than tending twenty live streams shaped like a casino floor. The wave design was for correctness. The batching turned out to matter as much for the person. Our guide to running and monitoring a workflow shows what a run looks like from the operator’s chair, and if the work spans several workflows, the World View puts them on one canvas instead of in seven windows.

The concrete next action costs one meeting: ask each person running agents how many they had live at once yesterday, and how many they could describe from memory. The gap between those two numbers is the attention deficit already on the books. It arrived quietly, it is entirely fixable, and unlike most of what lands on an executive’s desk in August, it can be fixed with a number and a sentence: this many at once, and every run ends in a decision.

The tools stopped asking for your attention this week. That is precisely why it is worth deciding, on purpose, where that attention now goes.