š§° Your process is the thing the agent needs, not another tool
Plus what a payroll story that went viral yesterday actually proves.
Three-minute read. The lead is Airtableās agent skills. Skip the last two sections if you are short.
PROCESS
Your process is the thing the agent needs, not another tool
What changed. Airtable is pushing agent skills: you capture the frameworks, approval paths and decision logic your team already uses, and hand them to an agent as a reusable workflow. Each agent carries instructions, tools, skills, memories, and rubrics it checks its own output against. The skills format came out of Anthropicās work on Claude and is spreading beyond it.
What it means at your size. The bottleneck was never the model. It is that your process lives in one personās head, in a shape nothing can execute. Writing it down so an agent can run it is the same work as writing it down so a new hire can run it, which is work you already owed yourself.
The first step. Take the one process you explain most often to new people. Write it as steps, inputs, and the rule for what ādoneā looks like. That document is the skill, whatever tool you eventually run it in.
The limitation. Airtable also reports a customer, Flatiron Health, whose non-technical ops team built in a month what a data engineer had scoped at a year. That is Airtableās account of its own customer, not an independent one, and a month of a teamās time is not free. Treat it as a direction, not a benchmark.
HANDOFFS
The hours are in the handoffs, not the tasks
Allie Miller argues companies are underinvesting in process management, and that the waste sits in the gaps between people rather than inside anyoneās work. Her example: the follow-up that should be drafted the moment a meeting ends, and instead waits on someone remembering. It is an argument, not a study, and the reason it is here is that it points at the cheapest thing you can fix this week. Look at where work stops moving and waits for a human, and automate the waiting, not the work.
Allie Miller on process management (x.com)
PACE
This will land slower than the noise suggests
Ethan Mollick expects AI to change professions more slowly than most predictions assume, because institutions get a vote. He points to OpenAI coordinating with mathematicians on how proofs get released, and expects law and medicine to negotiate similar terms through their bar associations and medical boards. If you are worried you are late, you are probably not. The constraint in front of you is your own process, not the pace of the field.
Ethan Mollick on institutional pace (x.com)
Just shipped
- Xiaomi MiMo-V2.6-Pro debuts as the top open-weights model on the Artificial Analysis index at 46, under an MIT licence, at $0.13 per benchmark task. One for whoever runs your automations. (artificialanalysis.ai)
- MiMo-V2.6-Flash, the cheaper sibling, 310B total and 15B active, for high-volume jobs. (artificialanalysis.ai)
- Jev is free on the Vercel AI Gateway until 25 September, and free via Lovable this week. After that, $0.042 per million input tokens with output free. (theunwindai.com)
The rest of the day
- Benchmarks are becoming the moat. Logan Kilpatrick says AI product teams should spend more than a quarter of their time building them. (x.com)
- Good training environments are scarce because checking the answers is the expensive part, per Nathan Lambert. (x.com)
- NVIDIA, NIPA and Artificial Analysis are building Korea a sovereign foundation model. (x.com)
- Harnesses are becoming routers. Lovable now routes between OpenAI and Anthropic depending on the job. (x.com)
Too early
Sovereign national foundation models. Several governments are now funding their own. Nothing here changes what you run this year, and the first versions will lag what you can already buy. Worth knowing the trend exists, not worth acting on. (x.com)
š One for the road
Xiaomi did not only release the weights. It released more than 7,000 reinforcement learning environments and the training code alongside them. The recipe, not the cake. That is the part the rest of the field will still be working through next year. (artificialanalysis.ai)