Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Verdict

Veeva's 180-customer rollout leaves queue verdict intact
08/27/26, 08:34 PM
Follows up on Veeva’s Workflow Agents Come With an Hourly Queue

Our Veeva workflow-agent analysis argued that the pilot should be judged by queue wait, exception rate, and reviewer minutes—not automated-task counts—because Vault 26R2’s hourly pickup could leave a task waiting nearly 60 minutes. Veeva now reports more than 180 Vault CRM customers live and 12 top-20 biopharmas committed, with one top-20 company deploying Vault CRM and Agentic Call Report across its full U.S. field team in the quarter. That is meaningful adoption evidence, but the company still does not publish realized queue times, production accuracy, exception rates, or net labor savings in this release. Verdict: the call holds—regulated teams should pilot bounded, nonurgent work, not expand on customer count alone.
Wire

Google Cloud adds hard caps to agent spend
08/27/26, 08:33 PM

Google Cloud added agent-spend controls that pause an agent’s API calls at a project-level monthly limit and send alerts at 50%, 80%, and 100% of budget. Its billing announcement also offers pay-as-you-go Gemini Enterprise consumption, pooled quotas, and Flexible Savings Plans that cut token costs 10% for one-year or 20% for three-year commitments; deferred execution pricing can be up to 50% cheaper but is marked for selected workloads coming soon. Teams putting agents into production should wire caps and anomaly alerts before broad access, then choose commitments only after usage stabilizes—the control plane is more durable than a vendor’s savings headline, a practical counterweight to the archive’s finding that only 6% of firms report real AI earnings impact.
Wire

Salesforce gives Claude 37 sales skills
08/27/26, 07:35 PM

Salesforce and Anthropic launched Claudeforce with Salesforce in Claude, a plugin containing 37 prebuilt sales skills for meeting preparation, deal-health review, pipeline analysis, and governed CRM updates. Salesforce’s announcement says select customers can pilot it now, with open beta expected in September; it also says Claude is the default model for Slack AI and reports 8.1 million annualized productivity hours inside Salesforce, a company-reported internal figure. For builders, this is a distribution test rather than a replacement verdict: wait for pilot pricing and evidence that permissioned CRM actions earn adoption beyond demos, while the archive’s enterprise-agent economics analysis supplies the comparison point.
Wire

AWS extends guardrails across three agent boundaries
08/27/26, 07:34 PM

AWS now shows 3 validation checkpoints for carrying Amazon Bedrock Guardrails beyond the model boundary: validate inbound context, inspect tool parameters before execution, and validate tool results before they reach a user or downstream agent. The AWS implementation wires those checks to Strands lifecycle hooks—BeforeInvocationEvent, BeforeToolCallEvent, and AfterToolCallEvent—and uses the ApplyGuardrail API; AWS also recommends deterministic schemas, regexes, and allowlists where latency matters. Builders using tools or MCP should treat model-only moderation as incomplete and add controls at the tool boundary, an approach that complements the archive’s analysis of agent evaluation and governance.
Wire

Gemini Omni 1.1 cuts video drafts 60%
08/27/26, 07:33 PM

Google’s Gemini Omni 1.1 Flash can generate 360p video previews up to 60% faster and at one-third the cost of its standard 720p output, while extending scenes in 10-second increments to a 40-second cumulative length. Google’s launch note also adds first-and-last-frame interpolation, up to three seconds of video references, and 1080p or 4K upscaling; the 60% figure is a system-throughput ceiling, not a quality-adjusted benchmark. For teams experimenting with video workflows, the practical filing rule is to prototype at 360p and reserve high-resolution renders for approved cuts, extending the archive’s analysis of generative video’s production economics.
Wire

NVIDIA posts $96.2B quarter
08/27/26, 06:33 PM

NVIDIA reported $96.2 billion in fiscal Q2 2027 revenue, up 106% year over year, while its Data Center segment reached $89.0 billion and management guided to $108 billion next quarter without assuming China data-center revenue. The company’s fiscal Q2 results put that forecast and geographic caveat in the same release. Builders should file the demand signal beside the $40M-per-megawatt Rubin deployment math: compute is still scaling faster than the market’s geography is settling.
Wire

Claude in Chrome goes generally available
08/27/26, 06:30 PM

Anthropic made Claude in Chrome generally available on every paid Claude plan, letting it navigate, type, click, and submit browser actions without asking for approval at every step. Its general-availability announcement says a safety classifier checks each action against the user’s request; the latest evaluation recorded 0% successful attacks against Claude Sonnet 5 and Opus 5 with probes plus classifiers, versus 0.3% against Fable 5. The practical implication is not to hand it production credentials blindly: keep reversible browser work in a controlled pilot, extending the queue-cost case for browser agents.
Wire

Socure reaches $5.2B valuation and buys Fravity
08/27/26, 06:27 PM

Socure announced a strategic growth investment at a $5.2 billion valuation and acquired Fravity, whose agent-building capabilities will be integrated into its RiskOS identity, fraud, risk, and compliance platform. The company announcement says RiskOS will gain first-party agent building for investigations, while the acquisition price remains undisclosed. For builders, the signal is that agent automation is being packaged inside an existing decisioning control plane rather than bought as a standalone chatbot; compare the archive’s benchmark on agent reliability before granting capital authority.
Wire

SWE Refactor Bench passes only 5.4%
08/27/26, 01:42 AM

SWE Refactor Bench found that only 28 of 520 runs (5.4%) completed a whole-repository stack migration, preserved behavior, and survived six independent agentic verifiers; 13 of 20 tasks received no accepted solution, according to the benchmark paper. The study’s three-stage funnel caught two opposite shortcuts: some agents skipped the migration while keeping tests green, while others migrated code and broke behavior. Teams should treat repository rewrites as gated migration projects—not one-shot coding-agent tickets—consistent with the archive’s evidence that the harness, not the model alone, is the enterprise moat.
Wire

Google puts legal agents in Gemini Enterprise
08/27/26, 01:39 AM

Google Cloud has put Gemini Enterprise for Legal into preview with reusable skills for contract review, regulatory scanning, legal research, and data-subject requests, plus permission-bound connectors to more than 10 legal systems, including iManage, NetDocuments, Everlaw, and RelativityOne, according to Google Cloud’s launch announcement. The platform is being developed with firms including Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly, and Google says customer data and outputs remain private rather than training its foundation models. Legal-operations builders should watch whether inherited matter permissions and ethical walls survive agentic execution, extending the archive’s analysis of legal AI’s enterprise valuation.