Agents run for hours now. Review is the product
Your agent does not need you anymore. That is the problem.
I wrote last month that autonomy needs a permission interface before it needs more power — approvals, audit, and clear boundaries. This week the data caught up. Agents now complete real engineering tasks without constant supervision, and the teams pulling ahead are the ones who built the review gate first.
Coding agents now run for hours on their own — and solo founders who ship treat the review, not the run, as the product.
What changed when agents started running for hours?
The unit of work got bigger. You used to get a snippet back. Now you get a branch.
SimScale's 2026 State of Engineering AI Report surveyed 350 engineering leaders and found teams with AI workflows generate nearly four times as many design variants per program. They also report a 2.8x speedup in servicing simulation requests. That is not a faster autocomplete. That is a different shape of work.
Here is my opinion, stated flatly: hours of agent output without a gate is not progress. It is a pile. Speed stopped being the constraint. Judgment is.
Why does trust decide what ships?
Because every team measured the same split: copilots everywhere, autonomy almost nowhere.
The same SimScale report puts copilot adoption at 67% in requirements engineering and 76% in simulation work, while only about 10% of teams deploy fully autonomous agents. And 75% of mature teams point to cloud-native infrastructure plus governance as the thing that lets them scale at all.
SimScale CEO David Heiny summed it up directly: "What we're seeing now is a shift from experimentation to scaled execution."
I read that as a warning. The teams running pilots doubled in a year. The ones who kept human oversight at the design gate are the ones seeing commercial return. The rest just generate more variants nobody trusts enough to ship.
Figma saw the same pattern from the design side. Figma's guide to agents designing directly on the canvas gives agents write access through the use_figma tool, but wraps it in skills and system components. Cat Wu from Anthropic put it well: "The best products come from teams who care deeply about the details." Access is cheap. Care is the filter.
How do you review an agent that works while you sleep?
You shrink the job before the agent starts, then read the result like an editor.
The plumbing matters here. The Model Context Protocol introduction for agent tool access describes the open standard most tools now share: one wire for files, tools, and context. Your agent can reach everything. That is exactly why the brief must be small.
Here is the setup I use on solo builds:
- One task per run. A single flow, a single fix, a single component. If the brief needs the word "and" twice, split it.
- Library only. The agent instances existing components and tokens. If it invents a new button, the run fails.
- A visible diff. Code changes come with tests. UI changes come with screenshots. No screenshot, no merge.
- A rollback line. Every approval names how to undo it. If you cannot undo it in minutes, you do not approve it at night.
Run that loop once on a real task from last week. If you would ship the result untouched, your system held. If you rewrote half of it, fix the brief before you blame the model.
When should you stop the agent and do it by hand?
When the question is still what, not how.
Naming a feature, cutting a screen, choosing the one action on a page, writing copy that carries money or legal weight — those are judgment calls. An agent will happily produce three options for each. All three will look competent. None will be yours.
My rule: if I am still deciding what the product says, I stay at the keyboard. Once I have decided and need it built in five places without drift, I hand the agent a tight brief and go make coffee. The agent handles the wiring. I keep the delete key.
Hours of autonomy changed nothing about who ships. Someone still has to care enough to say no.
Frequently asked questions
Agents moved from suggesting single lines to completing whole tasks without supervision. The output got bigger and faster, but the failure surface grew with it — one wrong assumption now compounds across dozens of files instead of one.
About the author
mosh
mosh is a product designer for growth, working with design thinking and ever-improving design systems. What matters: fixing conversion, whether in B2B dashboards or direct-consumer apps.
Keep reading
- Escape Ludo: game design
A covid quarantine project. Sprite design, character formation, levels, and inventory UI for a multiplayer Ludo game.
- Go Laundry: service booking platform
One platform to book cleaning services in Qatar. Art-directed brand mascots, app UI, and UX research.
- SaaS activity tracking: designing an admin log teams trust
Admins needed to see every team action in one place. Research with two managers and eight user stories turned a confusing log into filters, search, share, and export.