Featured
Signal2 min read
Fine-tune is usually the third move
Prompt and retrieval first. Train weights only after the behavior stays put and the evals show a plateau.
ReadArticles
Notes from the studio.
Short, high-signal writing on shipping AI and apps. Briefs, playbooks, and curated drops. Studio voice, not a personal diary.
Featured
The feed
Technique
Before you paste the new AI opt-out line, decide what AI may do with your store pages
Google, Microsoft and OpenAI say they will honor one shared robots.txt line for AI training. At Google, that line can also take your pages out of Gemini answers. Decide per page type first.
6 min read
ReadTechnique
Before Black Friday, check what AI agents say about your store
Shopify moved its agent Knowledge base into the admin, and your policies and contact details are already published to agents. Read what they will repeat, fix the gaps, and check where Google now shows ads next to those answers.
5 min read
ReadTechnique
Your agent can now pay a business by its @handle. Write the pay rule before you let it.
Stripe Pay sends business to business transfers to a Stripe Profile handle from the Dashboard or the Stripe CLI. Before an agent touches it, write down who it can pay, how much, from where, and who confirms.
4 min read
ReadTechnique
Check your cart hold rule before an agent games it (and before you block every agent)
The Verge used a shopping agent to keep The RealReal items on hold for hours by re-adding them every time the cart timer ran out. Test your own hold rule and limit what agents can do at cart and checkout.
5 min read
ReadTechnique
Make agents prove who sent them before your store says yes
A shopping agent can show up with a valid card and a working login and still have no proof that a person agreed to this order. Proof's x401 gives sites a way to ask. Try it on one route.
5 min read
ReadTechnique
Put your agent's monitor where the agent can't reach it
Outside researchers stress-tested OpenAI's internal agent monitoring and found ways around it. OpenAI's own reports show agents taking paths nobody authorized and keeping quiet. Here is a one-afternoon drill for your Always On Agents.
6 min read
ReadTechnique
Audit your agent's logins in 20 minutes, before your bank does it for you
Banks have started blocking AI agents that sign in with the owner's password, and Chase now puts agent errors on the customer. List every login your agent holds and move each one to the right kind of access.
4 min read
ReadTechnique
Write your store's agent access policy this week
BCG surveyed big retailers and found most have already picked a stance on outside shopping agents. Pick yours on purpose, draw the data line, and put it on one page.
5 min read
ReadTechnique
Draw your agent's money map before you connect it
Claude Duet launched with Shopify and Stripe on Oct 21, and the two handle refunds in opposite ways. Spend 30 minutes mapping every money action your agent can reach, then set the rules.
5 min read
ReadTechnique
AI agents are already probing your store. Run this drill
NorForge watched cheap AI agents pick e-commerce targets, check that orders worked, and push for a foothold on the server. Patch StyleSmuggler, list every script on your checkout, and check what your own Always On Agents can touch.
6 min read
ReadExplainer
What is an AI operating system for business?
An AI operating system is the set of Always On Agents, skills and routines that runs your recurring work in your own accounts, with you approving every spend, post and send.
5 min read
ReadTechnique
Any agent can now check out on your Shopify store. Buy from yourself first
Shopify's new Shop MCP lets any compatible agent search Catalog and finish a Shop Pay checkout once the shopper confirms. Run one test search and one refundable test buy on your own product this week, then write down whether you allow it.
6 min read
ReadTechnique
Send a shopping agent through your store before Muse's new model does
Meta's next Muse model is trained to try another route when a website stops it. Before it reaches your product pages, run real errands through an agent yourself and fix what it trips on, in six steps.
6 min read
ReadTechnique
Your agent pressed submit
Anthropic's test agents filled out real forms on real websites, including one they were told to stop before sending. Here is a submit gate and a 24-hour incident drill for Always On Agents, in seven steps.
6 min read
ReadTechnique
Give each Always On Agent its own seat, not your login
Google's new Gemini coworker agent gets its own account, its own audit trail and a spend cap that pauses it. You can set up your agents the same way today on the tools you already pay for. Five moves.
7 min read
ReadTechnique
Get ready for checkout inside TikTok before it reaches you
TikTok is testing one-click checkout from the For You feed and a shopping assistant that answers from your product data. Access runs through your sales rep, Buy Direct needs a UCP connection, and Salesforce's rollout starts in Q1 2027. Ask about eligibility, find your UCP path, fix the product data, and plan order reconciliation now.
6 min read
ReadTechnique
Measure a ChatGPT Ads test before you trust it
OpenAI named its first ChatGPT Ads measurement partners on Oct 5. Wire pixel plus CAPI with dedupe, connect the attribution tool you already run, set click and 1-day view windows on purpose, check Event Quality Score, then run a geo holdout before you trust in-platform ROAS.
6 min read
ReadTechnique
Decide who can message your customers before someone else does
On Oct 6 an unauthorised push notification reached ASOS app users through the third-party platforms ASOS uses to talk to customers. Your store can be fine while your messaging tools are not. List every tool that can send, move to named seats with 2FA and scoped keys, give agents draft rights only, and write the first-hour plan now.
5 min read
ReadTechnique
Define AI seat incrementality before the category does
Most AI seats still get judged on vanity activity. Publish an Outcome Spec, a baseline, and a client-verifiable method before anyone claims lift.
4 min read
ReadTechnique
Put a decision model on the agent hot path
Stop spending frontier tokens on triage, routing, and guardrail checks. Use Clef or Decider for typed answers; keep Sol/Sonnet-class models for generation.
6 min read
ReadTechnique
Migrate Gems to Skills before Argon opens
Google’s Workspace skills Rapid Release starts Mon 10/05. Inventory production Gems, rewrite as SKILL.md, create in both Gemini app and Workspace. Put Gemini 4 Argon on the routing watchlist at intro $2/$10. A/B only when paid API / AI Ultra opens.
6 min read
ReadTechnique
Sonnet 5.5 as the agency workhorse
Claude Sonnet 5.5 keeps Sonnet 5’s list price, cuts cost per task up to 30%, and runs 30%+ faster (Anthropic-reported). Claude Code already defaults Sonnet to 5.5. Migrate thinking, then flip traffic.
4 min read
ReadTechnique
GPT-6.1 Sol as the OpenAI workhorse
OpenAI’s DevDay upgrade nearly matches Astra on agentic coding and computer use at one-fifth Astra’s standard token prices (OpenAI-reported). Flip Codex and API agent traffic to gpt-6.1-sol this week; keep Astra for the hardest paths.
6 min read
ReadTechnique
Take the scoreboard away from the agent
Research agents that owned both the result and the evidence reward-hacked at a 30.5% rate on open-ended pipeline tasks. Move the metric, the scorer, and the tool grants outside the agent's reach.
6 min read
ReadTechnique
Deny-by-default before you give the agent a browser
Agents escalate instrumentally when blocked on ordinary data retrieval. Lock deny-by-default tool policy this week, then copy Anthropic’s measure→hill-climb loop with a human on every merge.
5 min read
ReadTechnique
Add runtime controls to your coding agent with NVIDIA OpenShell
OpenShell 0.1 puts a policy boundary outside Codex or Claude Code: deny by default, allow reads, route new access to a person. Here is the afternoon setup.
8 min read
ReadCurated drop
Friday drop: lean desk, openable work
Six picks from a thin week: openable DoD, shadow before write, desk not inbox, humanize before publish, cut costume labels, owner on the agent PR.
2 min read
ReadTechnique
Reprice your agent stack in one afternoon
Sep 21–22 stacked three frontier ships and a price cut. Pick this week’s default for billable agent work, then remeasure cost per task before your retainer still prices last month’s floors.
4 min read
ReadSignal
The coordinator just shipped
One AI/app shipping story: fleets got a boss. Here is what builders should do with it this week.
4 min read
ReadPlaybook
The agent seat is a control plane
Keep the agent seat for identity, permissions, approvals, and client context. Invoice a governed seat plus an Outcome Spec: setup, retainer, accepted work, and usage rails.
5 min read
ReadCurated drop
Friday drop: production AI without the theater
Five practices from the week: the verb, golden cases, an override, traces, and consent before the model call.
1 min read
ReadShip log
Ship log: 22atlas.com
What is live on our marketing site this week, what we threw out, and what is still ops.
3 min read
ReadPlaybook
TestFlight is not a launch
App Store Connect paperwork, invites, and AI consent are the release. The IPA is the easy part.
3 min read
ReadTechnique
Golden cases from last week's failures
Build the first eval set from real traces. Cluster the errors before you add another judge.
3 min read
ReadPlaybook
Finish the stalled build
Restart without a second greenfield. Name the stall, triage before git, lock one slice, finish a path someone can open.
4 min read
ReadTechnique
Shadow mode before the write
Run the model on real traffic without letting it touch the system of record. Compare, then promote.
4 min read
Read