The agents left the IDE

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. This week the coding agent stopped being something you run in a terminal and started being somewhere work happens: it got a chat channel, a file store, a skills registry — and its own git forge. Let's go.

1. The signal of the week

Agents moved out of the IDE and into the org chart — three launches in two days, same direction.

On August 20 Anthropic moved computer use, the Skills API, and the Files API to general availability, and added a browser use tool that clicks actual page elements instead of guessing coordinates on a screenshot. Computer use now takes several actions per model turn — Anthropic quotes a customer whose 32-minute workflow dropped to 13 (customer-reported, not a benchmark).

The same day Slack launched Slack Code: tag a coding agent — Claude Code, Devin, Copilot, or Vercel's — and it gets a dedicated channel with the plan, the diffs, and live previews, where the whole team watches and steers. A day later GitHub shipped its half: mention @GitHub in Slack or Teams and a Copilot cloud agent triages the issue, implements changes in a sandbox, and opens a PR — with optional human approval before merge (public preview, Business and Enterprise plans only).

Why this leads the issue: the pieces for building on this just went GA — state (Files), versioned logic (Skills), hands (browser and computer use for systems with no API), human checkpoints (approval gates). That's a parts list for a product that accepts a job and returns a finished artifact — much harder to replace with a ChatGPT button than another chat wrapper.

The caveat: GA means "supported," not "reliable on your workflow" — computer use is still the flakiest layer in the stack, and the most polished chat-agent experience shipped enterprise-first.

2. My take

Here's how I'd put it: enough chat wrappers around models. The far more interesting thing to build now is autonomous jobs — tasks that take input data and return a finished result.

Especially since nearly the whole toolkit for that is already here: Files for data and state, Skills for the logic, browser and computer use for services with no API. And anything critical can go through manual approval.

A product like that is much harder to make useless with the next model update. The user doesn't need yet another chat — they need a concrete result: a finished report, a processed document, a completed task.

But long chains are still fairly unreliable, especially against real interfaces through computer use. So I'd start trying this approach today — on one narrow, easily verifiable task. For customer-facing processes, treat it as prototype-only for now.

3. Tool worth trying

Qwen3.8-27B — with the reasoning turned down.

The launch was issue #010's story; this week brought the third-party verdict. Artificial Analysis scored it 52 on its Intelligence Index — #1 of 137 open-weights models in its class — and it passed 3 million Hugging Face downloads in three days. The field reports found the catch: the model ships with reasoning effort defaulted to xhigh, and at that setting Simon Willison watched "draw an SVG of a circle" burn 22,000+ reasoning tokens over about 21 minutes. Artificial Analysis logs the same pathology from its own runs: 160M tokens generated across its evals against a 45M median — "very verbose."

The scenario to try: pull the GGUF, then run the same ten tasks three times — reasoning off, low, and high. Measure wall-clock time, tokens, and pass rate. Use low or off for interactive work, high for overnight batch jobs.

I've been testing it since it landed — the video is on the channel: "Qwen 3.8 Locally: How Much Can You Actually Trust It?". My setup is a desktop with a 32 GB GPU; on my MacBook the model is too slow for real work. Treat "runs locally" as "runs on a real GPU."

The limits, stated plainly: even tuned, it's verbose; the index scores were measured at settings you may never run; and the vision claims still need field testing. But the headline stands — a genuinely frontier-class model now fits in a 17 GB file. You just have to configure it like you mean it.

4. MVP idea of the week

The headless job agent — pick one job, sell the artifact.

• Problem: small businesses have recurring document-shaped jobs — reconcile these invoices, prepare this filing, update listings across three portals — that have no API, so they're done by hand or not at all. Every "AI solution" they've seen is a chatbot that advises instead of doing.
• Solution: a single-purpose agent service: the user uploads documents → the agent applies your packaged expertise (a Skill) → uses browser use for the portal with no API → returns a finished artifact — a report, a filing, updated records — with a human approving anything irreversible.
• For whom: one niche you actually know, picked narrow — short-term-rental owners syncing listings, small importers reconciling customs paperwork.
• Minimal v1: one job, one Skill folder, the Files API for state, an email approval gate. No dashboard — email in, artifact out.
• Demand check: concierge first. Run five clients' jobs through your pipeline by hand-triggering it, charge real money, and only build the self-serve wrapper if they come back monthly.
• Why now: all three primitives — Skills, Files, browser use — went GA on August 20. This architecture was demo-grade last month; it's supportable now.

5. Useful links

• Anthropic — Build production agents with computer use, the Skills API, and the Files API — the GA post: the parts list for agents that live outside an IDE. Read it as an architecture diagram, not a product ad.
• Simon Willison — Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things — the week's most useful five-minute read: the overthinking pathology, the numbers, and the config fix. Read before your first ollama pull.
• Cursor — Origin code hosting — the week's other story: Cursor launched its own git forge "built for agent scale," in early beta on paid plans, on the same day GitHub spent ~6.5 hours degraded. Mirror mode keeps GitHub as source of truth — note the page says nothing yet about pricing or data terms.
• OpenRouter — OpenRouter is joining Stripe — the model router is being acquired (~$7.5B, per press reports). Read what "nothing about your integration changes" covers, then grep your projects for openrouter.ai and decide how thick your abstraction is.
• Anthropic — research on multi-agent systems — the "multiagent turf war" study: agents with conflicting goals deployed kill-loops and impersonation; the newest models mostly negotiated truces, earlier ones fought. Read before wiring two agents into one repo.

Which side of the week are you on — tuning a local model's reasoning down, or giving a cloud agent its own chat handle? Reply with one line; it steers what I test next.
— Nick

P.S. One more thing — I launched a Patreon. Every video on the channel is the polished end of a longer process: models tested, tools broken, opinions formed and sometimes reversed. The Patreon is that process, published as it happens — essays, AI news analysis, reactions to releases, and the code and configs from my videos and projects. To be clear: this newsletter, the channel, and my open-source stay free — nothing moves behind a paywall; the Patreon runs ahead of them. $10/month, 4–5 posts and videos a month guaranteed, no hype there either. Three posts are already up: my reaction to Monday's GitHub outage is open to everyone, the other two are patrons-only. Read the open one and decide if the format is for you: patreon.com/ZerotoMVP