|
Hey there 👋
For two years, “an agent that does your knowledge work end to end” was one product with one name. This week it became a category. Two days after Claude Cowork left the laptop for your phone and your inbox, OpenAI shipped its answer — ChatGPT Work — and made GPT-5.6 generally available underneath it.
Here’s the Friday thought worth carrying into the weekend: when a thing goes from one product to two, the asset you build isn’t the choice between them. It’s the contract they both have to honor. More on that below.
The Big Thing
OpenAI ships ChatGPT Work — the coworker-agent is now a two-horse race
OpenAI launched ChatGPT Work yesterday, and it’s the clearest sign yet that the general-purpose “coworker” agent is a real product category, not one company’s bet.
What it does: you hand it an outcome, and it works across your apps for hours — breaking the job into steps and finishing them on its own. The output isn’t a chat transcript. It’s finished material: spreadsheets, slides, documents, even interactive web apps. Codex is built in, and it runs on GPT-5.6, which OpenAI made generally available the same day across ChatGPT, Codex, and the API (Codex itself is now folded into the ChatGPT desktop app).
The part that matters for how you deploy it: the control surfaces are built in. There’s Plan mode — a step-by-step plan you approve before any work starts — plus configurable check-ins and explicit action approvals. Press framed the whole thing as OpenAI’s direct answer to Anthropic’s Claude Cowork, and that framing is the point.
Because here’s the third-thought. The two leading coworker-agents now ship the same primitives: a plan you approve, check-ins along the way, sign-off on actions, a trail of what happened. The model tier underneath (GPT-5.6 comes as Sol flagship, Terra balanced, Luna cheap-and-fast) is a knob, not a moat. So the thing you should actually design around isn’t which coworker wins. It’s your own workflow contract — the one both of them have to obey.
Ship it? Pilot — don’t standardize on one. If you already run ChatGPT Enterprise or Codex, ChatGPT Work is a same-account pilot this month. But write the contract first: which steps require Plan-mode approval, which actions need a human sign-off, what gets logged. Make it identical to what you’d demand of Cowork — so the coworker stays swappable. On the model, route routine turns to Luna or Terra and escalate to Sol only when the task earns it.
Sources: 9to5Mac, US News/Reuters, CryptoBriefing (Jul 9). “Answer to Cowork” is press framing, not an OpenAI claim.
Tour de Headlines
🧩 Microsoft’s Sales Agent and Service Agent hit GA — inside the tools reps already use. Both went generally available (announced Jul 7), living in M365 Copilot, Outlook, Teams, and Dynamics 365. Sales Agent writes meeting takeaways, objections, and commitments straight into CRM fields without an app switch; Service Agent summarizes case context and pulls answers across Dataverse and SharePoint. Both are grounded in live CRM data over an MCP foundation. The non-obvious read: the coworker fight isn’t only the standalone super-apps — incumbents are shipping the same capability where your team already sits. (Work IQ, the layer underneath, was our #8 Big Thing; the agents going GA is the new part.)
🛡️ DeepMind and partners put $10M behind multi-agent safety. Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, and ARIA opened a funding call (up to $10M, deadline Aug 8) for research into how populations of agents behave — the risks that emerge when agents interact, not when one model acts alone. Four clusters: sandboxes, the science of agent networks, agent infrastructure, and multi-agent oversight. It’s the Friday counterweight to a capability-heavy week: as you wire up two coworkers, CRM agents, and data-writing agents, the open question is what they do to each other.
⚖️ Norm Ai raised $120M at a $1.2B valuation — for the agent that watches your agents. The round (Khosla-led; Blackstone, Bain Capital Ventures, Coatue, Vanguard, New York Life, and TIAA in) funds agentic legal and regulatory work. The tell isn’t the unicorn tag. It’s that Norm is building agents to monitor other AI in regulated industries as a compliance check — the money is flowing to the supervisor layer. Clients are cited at >$30T in assets under management (their figure).
|
Sponsor
You’re writing an approval contract for your AI coworker. Your revenue calls still have no gate.
Today’s whole issue is one move: don’t just pick the coworker — define the contract it has to honor. Fair. Now point that same lens at the highest-stakes system in your business that still has no gauge: the live sales and CS conversation. The discovery call that got misread. The deal where two people “aligned” on different things. The rapport that never formed — and closed-lost with no warning light. RapportScore scores how your team actually connects, on every call and email, so the misfires surface while you can still fix them. You wouldn’t let an agent write to your CRM with no approval step. Stop running revenue conversations blind.
See your team’s score →
|
Tool of the Day
✍️ Airbyte Agents (write ops)
For teams whose agents can read the CRM but can’t touch it — now they can write, with a CLI and a marketplace listing.
Airbyte closed the gap between “agent that knows” and “agent that does” on Jul 8. Airbyte Agents landed in the OpenAI App Marketplace, and the Salesforce and HubSpot connectors gained write operations — agents can now update records, change deal status, and automate workflows. Flip a deal stage from ChatGPT and watch it reflect in Salesforce. A new single-binary Agent CLI drops the whole thing into terminals, scripts, and CI/CD, and it speaks MCP for Claude, ChatGPT, and Cursor.
It’s the “data access layer → operational execution layer” jump — the same write-access move that made last week’s coworker launches matter. Deploy note: turn write on for one object (say, Opportunity stage) behind an approval step before you widen scope. Same plan → approve → log contract as the Big Thing.
See Airbyte Agents →
Worth a Click
- GPT-5.6’s price tiers landed — Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per 1M in/out tokens. Re-run your routing math: cheap default, escalate to Sol only when the task earns it.
- Akeneo’s Agentic Ziggy is a clean, small example of the build pattern worth copying — a supervisor agent routing narrow specialists (modeling, mapping, enrichment, QA) beats one do-everything agent.
- Microsoft Service Agent GA detail — Dataverse + SharePoint grounding for support teams, if you’re weighing the in-CRM route over a standalone coworker.
The coworker-agent stopped being a product this week and became a category — ChatGPT Work and Claude Cowork now ship the same control primitives: a plan you approve, check-ins, action approvals, a log. So the weekend build isn’t picking one. It’s writing the one-page contract the coworker has to honor — what it plans, what you sign off on, what it’s allowed to write, what it logs — and making it engine-agnostic. Own the contract; rent the coworker.
Have a build-something weekend.
— Ace, for The Agent Stack
|
