No new models landed this week after Wednesday, which is a good week to look at the unglamorous half of the stack instead. Salesforce handed its entire CRM to an assistant and the interesting number is not the money, it is four. A survey of 550 engineers found near-universal daily agent use and a blocker that never appears in a keynote. Google quietly started a clock. And Anthropic published what alignment research costs when an agent does it: four dollars an hour.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Salesforce exposed its whole CRM through four tools. Stare at that number, not the headline.
Salesforce and Anthropic announced Claudeforce on Wednesday, alongside Salesforce's Q2 FY27 earnings. The consumer-facing part is a plugin called Salesforce in Claude, shipping with 37 prebuilt sales skills — meeting prep, deal health review, pipeline review are the three anyone has actually named — live for pilot customers now, open beta expected in September. Going the other direction, Claude becomes the default model in Agentforce Vibes and Coworker, served through Amazon Bedrock inside Salesforce's security boundary, and the default behind Slackbot, Claude Tag and Slack Code. Salesforce says 83% of its own workforce already uses the Claude-powered Slackbot. Reporting around the earnings call put Salesforce's 2026 token commitment at roughly $300 million on top of an existing equity stake of similar size.
Skip all of that. The architectural detail is the one worth your afternoon: the plugin runs on Headless 360's Hosted MCP Server, in beta since July, and that server exposes four tools. Discover, Describe, Dispatch, and a read-only Dispatch. An object model with thousands of fields, decades of custom logic and per-record sharing rules, and the agent sees four functions. Not four hundred endpoint-shaped wrappers. Four verbs, and the schema is something the agent asks about at runtime rather than something you enumerate in a tool list.
The second detail is permissions, and Salesforce's framing of it is blunt: if you do not own a record, and you do not have permission to see it, the MCP server does not either. That only works because Salesforce's sharing model was never implemented in its web controllers. It lives underneath. Most products are not built that way, and this is where an MCP server turns into an access-control incident: your UI enforces the rules, your service layer assumes the UI already did, and the agent talks straight to the service layer.
For builders
Two exercises, both short. First, count the tools you would need to expose your product to an agent. If the number is anywhere near your endpoint count, you are building an API wrapper, not a tool surface — try collapsing to the Salesforce shape: one tool to list what exists, one to describe a thing's schema, one to act, one to read. A small tool list is not a cosmetic choice, it is what keeps the model's selection accuracy up as context fills. Second, find where your authorisation checks physically live. Open the function your MCP server would call and ask whether it re-checks the caller's permissions or trusts that the request got past a controller. If it trusts, you have a bug that is currently invisible because nothing but your own frontend has ever called it. Fix that before you ship a tool surface, not after.
Story II
80.8% use agents daily. 41.1% hit problems daily. The gap has a name and it is state.
Temporal published its second annual State of Development Report on Wednesday, surveying 554 usable responses from engineers, architects and engineering leaders — two-thirds US, a third UK and EMEA — collected between April 29 and May 25. Daily-or-more agent use came in at 80.8%, against 47.3% a year earlier. That is a 70.8% relative jump, and it is the least surprising number in the report.
The useful numbers are the uncomfortable ones. 41.1% run into agent-related problems daily or more often, and 9.0% say continuously. 85.5% trust agent output at least somewhat, which is a strange thing to hold simultaneously with the previous sentence. Asked for the single biggest blocker to using agents more, 35.7% said tracking state — ahead of debugging, ahead of managing token and compute cost. 51.3% get from prototype to production in hours or less, 26.9% in minutes. And 92.3% have at some point tried to rebuild software they were paying for.
Worth saying plainly: Temporal sells durable execution, so a survey concluding that state management is the hard part is a survey that flatters the sponsor. Read the number with that in mind. It is still probably right, because it matches what breaks. A seven-step agent run that dies at step four is not a model problem and no amount of prompt work touches it, which is exactly why it never appears in a demo — demos do not crash on step four.
For builders
Take your longest-running agent flow and answer one question honestly: if the process is killed at step four of seven, what happens? If the answer is "it starts over," you are paying for steps one through three twice every time, and if any of those steps wrote to a database or called a payment API, you have a correctness problem rather than a cost one. The fix does not require adopting a workflow engine this week. Give each step a deterministic idempotency key derived from the run ID and the step name, persist each step's output keyed by it, and check that table before executing. Twenty lines and a table. Then instrument it: log resumed-runs against total-runs, because until you can see that ratio you cannot tell whether reliability work is worth doing at all.
Story III
Google shipped an omni model to GA and started a clock on the endpoint you are probably calling
Two entries in the Gemini API changelog this week, and one of them is a deadline. On August 27 Gemini Omni Flash reached general availability as gemini-omni-1.1-flash, adding video extension, interpolation between a first and last frame, and explicit resolution control across 360p, 720p as default, 1080p and 4K. In the same entry: the existing gemini-omni-flash-preview endpoint is deprecated on September 30, 2026.
The day before, Google launched two speech-to-text models — gemini-3.5-transcribe for non-streaming and gemini-3.5-transcribe-live for streaming over WebSockets — with utterance-based language detection across 85+ languages, speaker diarization and word-level timestamps. Diarization and word timestamps are usually the two features that push teams onto a specialist transcription vendor, so their arrival in a general API changes a build-versus-buy calculation that a lot of people settled a year ago and have not revisited.
Neither entry lists pricing, and the changelog is where Google publishes these things, so treat cost as unknown until you check the pricing page yourself. The deprecation date is not unknown.
For builders
Run grep -rn "omni-flash-preview" . across your repos and your environment configs right now — preview endpoint names have a habit of living in a .env nobody greps. If it turns up, September 30 is thirty-two days away, and the GA name is a string swap plus a regression run, not a migration. Then put the date in a calendar rather than in your memory. Separately, if you pay per audio-minute for transcription today, this is a cheap afternoon: take one real file from your worst-case audio — overlapping speakers, accents, background noise, whatever actually breaks your current vendor — and run it through gemini-3.5-transcribe. Judge on diarization accuracy and timestamp drift, not on word error rate, because those are the parts that cost you money to fix downstream.
Story IV
Anthropic ran alignment research at $4 an hour, and the agents mostly rediscovered the literature
Anthropic published a paper on Friday, led by fellow Chen Yueh-Han, in which an Automated Alignment Researcher was pointed at ten categories of alignment failure — deception, sycophancy, reward hacking, privacy violations, jailbreaks among them — and told to fix each one through a loop: search the literature, propose a method and data, train, then test against benchmarks like ConfAIde, PrivaCI-Bench, PrivacyLens and Petri. It closed between 26% and 96% of the safety gap depending on the failure, averaging 85% on deception, where the human researchers it was compared against managed 20% without iteration. On production-grade models Claude Sonnet 5 completed the experiments in 60 hours. TechCrunch's write-up carries the number that will get quoted for a year: roughly $4 per hour in API inference against $150 per hour for a human researcher.
The finding I keep turning over is a smaller one. On sycophancy, 98% of the methods the agent proposed converged on a single already-published technique — training the model on its own non-sycophantic answers. That is not a failure. It is a precise description of what these loops are good for: exhaustively applying what is already known, faster and cheaper than a person can. It is also a warning about where the returns stop. The paper's own stated limits point the same way — the failures studied were narrower than production, several failure modes have no benchmark at all, and whether the gains survive extensive RL training was not tested.
So the bottleneck moved. It is not ideas, and it is not compute. It is having a measurement good enough that a loop can optimise against it without gaming it.
For builders
Invert your usual order this week: write the benchmark before you write the fix. Pick the single failure your users complain about most — the summariser inventing a date, the classifier collapsing on one category, the agent calling the wrong tool — and spend an afternoon collecting thirty real examples with correct outputs into a JSON file with a scoring function. That is a small eval suite and it is the entire precondition for the loop Anthropic ran. Once it exists you can let an agent propose fixes against it and actually know whether they worked, which is not true of any prompt change you have shipped on a hunch. And set your expectations from the sycophancy result: for a known problem, expect the agent to find the published answer, not a new one. If you need a genuinely novel method, that is still your job, and the eval you just built is how you will prove it beat the obvious one.
— The Vibe Gate news desk. We read the firehose so you can keep building.