|
Hey there 👋
I spent this morning reading three launch posts from three different labs, waiting for the part where someone announced a model. It never came. OpenAI, Google, and Anthropic all shipped in the same stretch, and not one of the headline items was a new model. They were products for running the agents you already have: a governed platform, a background-task API, and an assistant that keeps working after you close the laptop.
So here is my bias, for what it is worth. The model stopped being the hard part months ago. The hard part is getting an agent to survive contact with production, and this week the majors admitted it by selling the deployment instead of the model.
The Big Thing
OpenAI will send engineers to run your agents.
Last week OpenAI launched Presence, a managed enterprise platform for building, deploying, and operating governed voice and chat agents. Read past the demo and the shape is unusual. It is limited general availability, not self-serve. OpenAI Forward Deployed Engineers and a few systems integrators lead the rollout: they pick the workflows, connect your internal systems, set the permissions, configure the policies, test the agent, and move it to production with you.
That is a consulting-shaped product wrapped around an operational layer, not an API you wire up over a weekend. Presence packages the parts nobody demos: policies and permissions, simulations, evaluations, human approvals, and a controlled update process. Codex plus a Presence plugin reads production signals, suggests changes, and lets you test each one against the version that is live before you roll it out.
The tell is in the dogfood. OpenAI runs its own English-language phone support on Presence. Its numbers: the system resolves 75% of inbound calls without a human and cut handoffs by 15 percentage points in 10 days. That is OpenAI grading OpenAI on OpenAI's own support line, so weigh the percentage accordingly. Believe the mechanism, which is a frontier lab running its own governed platform on real, account-verified support calls. Early named customers are BBVA Mexico, SoftBank, and IAG in Australia.
So what does it mean if you are not a bank with a forward-deployed engineer on speed dial? You probably cannot buy Presence yet. You can copy its shape for free. Define the policies and permissions before the model. Put every high-stakes action behind a human approval. Evaluate the agent against production traffic before you trust it. Keep a person in the loop on the calls that can lose money or a customer.
Ship it? Watch the product, steal the pattern. The thing every one of these launches quietly reveals is that the moat is the governance layer, not the model behind it.
OpenAI Presence · VentureBeat
Tour de Headlines
🧩 Google gave Gemini agents a background thread.
Google expanded Managed Agents in the Gemini API, and the useful bit is boring in the best way. Set background: true on a long-running task and poll for it asynchronously instead of holding an HTTP socket open and praying it does not drop. You can register remote MCP servers or define custom functions to reach your own APIs and databases, and credentials now refresh across interactions. One endpoint, and Gemini handles the reasoning, code execution, package installs, file management, and web access inside an isolated cloud sandbox. The fragile part of a twenty-minute agent run was always the open connection. This turns durable, multi-step jobs into a normal API call. Ship it? Deploy if you already build on Gemini. It is the smallest change with the biggest reliability payoff this week.
Google: Expanding Managed Agents in the Gemini API
📱 Claude Cowork follows you off the desktop.
Anthropic is rolling Cowork out to web and mobile, beta first for Max users. The line that matters for builders: work continues in the background, and scheduled tasks now run with no device online, so an agent job does not need your laptop awake to finish. Chat and Cowork share one home on web and desktop now, and voice mode runs on Opus, Sonnet, and Haiku, reaches your connected tools, and speaks more languages. Anthropic doubled Cowork usage limits through August 5 to mark it. Ship it? Try it if you already run Claude. Kick off a scheduled task, close every device you own, and confirm it lands.
Claude: Cowork on web and mobile
🛡️ Qoder moves security inside the coding session.
Qoder, an agentic coding platform that reports 5 million-plus registered users, launched Qoder Security: three progressive review layers that run as code is written and fix issues in the same session, before commit. It is built on Qoder's own security model and, per the company, checks whether a flagged issue is reachable in the real call path via taint propagation before it bothers you, then drives the fix itself. Qoder's reported figures: roughly 60% better detection, about 80% fewer false positives, discovery-to-fix from days to minutes, and 600-plus issues surfaced in internal pre-launch testing. Those are the vendor's numbers on the vendor's benchmark, so treat them as a claim, not a result. Ship it? Watch. Promising if you ship with an agentic coder, but wait for outside numbers before you retire your existing scanner.
Qoder Security announcement
|
Sponsor
Measure the human signal in every conversation.
Your agents get scored on every call they handle. RapportScore does the same for your people, reading the communication signals in real sales and support conversations so teams can see what lands and adjust before the next one. Built by the crew behind this newsletter.
See your team's score →
|
Tool of the Day
🧩 Gemini background agents (background: true)
What it is for: kick off a long agent job and stop babysitting the socket.
If you build on the Gemini API, this is the concrete thing to try this week. Set background: true on a Managed Agents call, get a handle back, and poll status async instead of blocking on an open connection. Register your internal tools as a remote MCP server or wire them in as custom functions, and let Gemini's sandbox handle the code execution and package installs. Start with one batch job you already run on a cron. Ship it? Try it on a non-critical, long-running task first, prove the polling and the failure path, then move something that matters.
Read the Gemini update
Worth a Click
- OpenAI Presence, primary. What the platform really packages, in OpenAI's own words. OpenAI post
- A deployment problem, not a platform problem. Why most enterprise "agents" are still chatbots that never reached production. VentureBeat
- Managed Agents in the Gemini API. Background tasks, remote MCP, and function calling. Google post
The loud story this week was protocol and security: MCP going stable, kill-switch bills, kernel sandboxes. The quiet story is bigger. Three of the majors shipped managed agent products in the same stretch, and every one of them buried the model. OpenAI is selling forward-deployed engineers. Google is running your agent's background job in its sandbox. Anthropic keeps the job alive after you shut the laptop. The center of gravity moved from "here is a model" to "here is a governed thing that survives production."
If I had to bet, the teams that win the next year are not the ones with the best model access. They are the ones with the best deployment story: policies, evals, a human on the risky calls, and a rollback that works. Steal that whether or not you buy any of these.
Ron
|
