An AI job can keep running after you close your laptop. So can the bill.
That connection kept coming up in Koda's coverage this week. Agents are taking on longer jobs. Software companies are finding ways to charge for that work. And the people buying it still need to know what happened, what it cost and who can approve the result.
I think the useful unit to watch is the finished job. Four stories help explain why.
Watch Koda This Week · 7 min 24 sec
The job outlives the chat
Cursor's cloud agents run independently of your computer's connection. That is an existing capability we explored this week, rather than a new launch. It changes the handoff: you can leave a job running and return to inspect the result.
Anthropic offered another view. In its own assessment of August AI research and development work, it classified 26% 03 as AI-led. Humans still supervise in that category, and Anthropic used models to help evaluate the work. It is not a measure of payroll savings.
My interest is in the handoff itself. How much of the job can an agent carry through, and what evidence does the person reviewing it get back?
Sources: Cursor cloud agents · Anthropic's measurement and limits.
The bill follows the work
Our Salesforce coverage needs one clear qualification: Salesforce still offers per-user licenses alongside consumption options for Agentforce. Seats have not disappeared.
The buyer's calculation does get harder. Cursor's cloud agents are billed at the selected model's API pricing. Work can generate usage while the number of people stays unchanged.
I would want to see the cost of a completed, reviewed job before celebrating a smaller seat count. Include the time spent correcting the result. A cheaper attempt can still be an expensive way to finish.
Sources: Salesforce's pricing options.
Three questions to bring to the next trial
What counts as finished?
Define the completed result before comparing tools.
What did the reviewed job cost?
Count usage and the time spent correcting the result.
Who can approve the result?
Make access, review and responsibility explicit.
Permission is part of the product
An agent needs access to do useful work. Reading a customer record and changing it are different permissions. Drafting a payment recommendation and approving it are different responsibilities.
Palantir's AIP documentation describes granular permissions, audit trails and approval workflows. That is the grounded point from our governance coverage; it does not establish a blanket internal ban on AI.
This is where the work and the bill meet accountability. If an agent spends money or changes a record, somebody needs to be able to explain what it was allowed to do and what actually happened.
Sources: Palantir's governance controls.
The model providers want more of the job
Anthropic introduced Claude Docs and Slides on September 16, in beta on paid plans, with Enterprise admins controlling access. OpenAI announced Astra for Law the next day, initially for selected law firms, with specialist partners part of the plan.
These announcements bring more work into the model provider's product. They do not tell us which specialist businesses will win or disappear.
My read: a useful feature becomes harder to defend when your supplier offers it too. A dependable business process still needs customer context, integrations, approvals and someone responsible for the result. I would look there for lasting value, while watching how far the providers move.
Sources: Claude Docs and Slides announcement · Astra for Law announcement.
What I would take into next week
The question I'm carrying into next week: how much of a real job can these systems finish reliably, at a cost we understand, with responsibility we can explain?
Pick one recurring task in your own work. Before trying another tool, write down what a finished result would look like, who would check it and what you would need to measure. That gives the next announcement something useful to answer. Some jobs will still need too much supervision. Record that too, so the next trial starts with what you learned.
Arno · Koda Intelligence
The editions behind this recap
- Koda: continuing work · September 13
- Koda: applications and specialist value · September 18
- Koda: the research measurement · September 19
Coverage and production notes
Koda editions published September 13-19, as observed on September 19. The final day is partial. Edition dates are not necessarily product launch dates.
Coverage cutoff: September 19, 2026 at 4:04 p.m. America/Chicago. This is an additional weekly Deep Dive.
AI-assisted writing and editing. The video uses Arno's authorized ElevenLabs voice clone and an AI-generated presenter based on his photograph, alongside original Koda excerpts and generated illustrations. The presenter is not a live camera recording. The connections between stories are editorial interpretation.