The Intel Report
AI labs and platforms exposed more of the numbers underneath delegated work: how much research AI performs, how agent activity is monitored, which features people use, and where higher-risk access begins and ends. The metrics are still vendor-defined, but they give operators something concrete to test.
Reporting window: September 12-18, 2026 · Now You're Technical
Published September 18, 2026
Anthropic published measurements for AI-led research, agent oversight, and safety compute. GitHub added feature-level and customization-level usage data. Google moved agent review into traces, runtime policies, and follow-on enforcement. A new life-sciences access program tied stronger capability to identity, purpose, renewal, and monitoring. The same week also produced reusable scientific code, an open SDK-generation stack, a short agent-runtime migration clock, and models designed for narrower operating jobs. My read: the work is becoming governable because it is becoming observable. Vendor dashboards still need local outcome measures, but they are a better starting point than a capability claim alone.
Anthropic's proposed measurements expose how much work AI performs, how its actions are reviewed, and where compute goes. They are self-reported prototypes, not cross-lab benchmarks.
Anthropic's August snapshot says Claude led 26% of measured AI R&D work, more than 90% was at least AI-collaborative, and none was fully autonomous. The index uses a fixed task tree, model-generated evidence, a model judge, and person-time weighting.
Read the original source →On Anthropic's most-used internal platform, roughly 30,000 research and engineering agents were active at a time. Anthropic says every action passed through online monitoring and entered offline monitoring; about 1 in 47,000 August decisions was blocked, and roughly 50 high-priority flags reached people each week.
Read the original source →Anthropic classified about 6% of AI R&D compute as safety work and about 12% of AI-driven AI R&D compute as safety work in one measured week. The company says the categories rely on best-effort metadata and a prompted classifier, and that compute is an imperfect proxy for effort.
Read the original source →Operator read: Ask for the denominator, method, coverage boundary, review latency, and independent check behind every governance number. A dashboard can show activity without proving that the control works.
GitHub's new Copilot reporting shows which features and agent customizations enter regular work. The detail is useful, but usage is still one step away from business value.
The Copilot impact dashboard now counts active users who engaged with an included feature on at least two days in a rolling 28-day period. The same aggregate view is available through enterprise and organization APIs.
Read the original source →Reports break engagement out across code completion, agent edit, active and passive code review, cloud agent, CLI, and the Copilot app. A person can appear under more than one feature, so the categories should not be added into one user total.
Read the original source →Copilot CLI reports now include top-five and distinct-use measures for skills, custom agents, MCP servers, slash commands, and plugins. One important limit: the MCP interaction count records connection attempts, including failures, rather than each tool call.
Read the original source →Operator read: Measure adoption at the workflow level. Pair feature use with accepted output, reviewer corrections, cycle time, and a business result. A frequently used agent can still produce expensive rework.
Google's private preview and reference implementation treat behavior across a session as an operating record that security teams can review and act on.
Agent Anomaly Detection reads logs, OpenTelemetry traces, reasoning traces, tool calls, and execution flow for agents on Gemini Enterprise Agent Platform. The service analyzes them asynchronously outside the live request path.
Read the original source →Findings carry severity, rationale, and recommended action, and also appear in Security Command Center. An API lets a callback or plugin block later tool calls or halt a later turn when a threshold is crossed. That is follow-on enforcement, not proof that the first harmful action was prevented.
Read the original source →Google's zero-trust pattern combines Model Armor for prompt and response screening, semantic governance policies for intent-aware tool decisions, and anomaly detection for behavior across multiple turns.
Read the original source →The platform enforces policies separately from the application, so a security administrator can change a tool constraint without rebuilding the agent. The companion demo shows how a series of individually allowed refunds can become an anomalous session; its confidence values and finding payloads are illustrative.
Read the original source →Operator read: Define which traces exist, how long they are retained, what the detectors miss, who reviews a flag, and how quickly a policy change takes effect. Observability becomes a control only when someone owns the response.
Anthropic's life-sciences beta grants stronger biology access through verification, declared purpose, renewal, monitoring, and explicit product limits.
Standard Use applies across a verified team and renews annually. High-risk Use applies to a specific project and renews every six months. It removes safeguards that block life-sciences requests while leaving other controls, including cyber classifiers, in place.
Read the original source →Applicants are reviewed for research credentials, security standards, and ethical oversight. Access is tied to stated use cases, and the program shifts some enforcement from real-time blocking to offline review of behavior across requests and sessions.
Read the original source →LSVP retains associated data for 30 days so flagged activity can be reviewed. At launch it is unavailable to BAA-enabled organizations, individual plans, and third-party platforms. Anthropic directs customers with PHI to separate non-BAA, non-HIPAA organizations.
Read the original source →Operator read: Treat elevated model access like a time-bounded entitlement. Tie it to a person or team, a stated purpose, an expiration date, monitoring, and an incident path before the stronger capability reaches production work.
Anthropic's biomolecular work is useful because it produced code, benchmarks, and visible failure cases. The results remain vendor-authored and need independent reproduction.
The report says Claude produced optimized packages for 36 implementations covering more than 30 open-source biomolecular models in just under four weeks. Two domain scientists without prior inference-optimization or kernel-engineering experience provided most of the supervision.
Read the technical report →On H100 GPUs, Exact mode averaged a 1.6x forward-pass speedup across 14 structure-prediction models; Fast mode averaged 4.1x across 13. Across 1,925 model-target pairs, pooled acceptable-interface rates changed by less than one percentage point and no pooled change was distinguishable from zero.
Read the technical report →Big mode accurately modeled several systems above 10,000 tokens on one eight-GPU node. It also completed inference up to 70,320 residues on one eight-B300 node, but the largest predictions collapsed and were inaccurate. Anthropic released the optimized code for inspection.
Read the research announcement →Operator read: Prefer work that leaves behind a reusable artifact and a reproducible test. Have a domain specialist inspect the benchmark, run the code on representative data, and check the complete workflow before changing a production pipeline.
One supplier shutdown and one short preview retirement show how quickly an agent workflow can inherit risk from generators, schemas, and runtime contracts.
Google says its proprietary SDK-generation supplier was acquired and abruptly shut down in May, during the run-up to Google I/O. The replacement now powers client generation for the Interactions, Agents, and Webhooks APIs.
Read the original source →Speakeasy is open-sourcing generators for SDKs in seven languages, standalone agent-native CLIs, and documentation MCP servers. Google says generated clients may keep a chosen license; changes to the generator remain subject to AGPL, which deserves local legal review.
Read the original source →antigravity-preview-09-2026 replaces the May preview, which shuts down October 5. Local tool users face PascalCase parameters, line-range file replacement, and new built-in search tools. Remote consumers that only read output steps may need only the new agent identifier.
Operator read: Put interface contracts on the risk register. Pin versions, test tool schemas, rehearse a provider retirement, and decide which client-generation layer your team can keep running if a supplier disappears.
Google separated live audio by latency and reasoning needs. TypeSafe went further and removed free-form generation from a model designed for typed decisions.
Google made gemini-3.8-live and gemini-3.8-live-extended-thinking generally available. The first targets low-latency dialogue with interleaved reasoning and asynchronous function calls; the second adds background reasoning for more demanding audio interactions.
TypeSafe released Jev in early access. The service accepts application state and predefined typed questions, then returns structured choices with probabilities and confidence rather than generating a free-form response token by token.
Read the original source →TypeSafe lists 70-500 ms response times, $0.042 per million input tokens, and no output-token charge. Its headline speed and cost gains come from provider-run workflows, use Astra and Fable as reference probabilities, and are described by the company as the high end of expected real-world gains.
Read the original source →Operator read: Route by the job. Choose low latency for conversation, deeper reasoning when delay is acceptable, and typed decisions for narrow program branches. Benchmark each choice on your error cost and failure handling, not the provider's best case.
Make it practical
Evidence and limits
This edition draws on 11 first-party publications, technical reports, and dated changelog entries. Multiple cards may draw distinct findings from the same source; 22 signals does not mean 22 independent studies. Vendor measurements, early-access performance, and private-preview behavior are attributed claims. Operator reads are our analysis.
All material claims were checked against the original sources linked below. The source monitor checked 43 lanes and produced 10 current captures, with five HTTP 403 gaps. Scout and the preserved podcast corpus had no current-window artifacts. X API records and current podcast or newsletter captures informed discovery only. A same-day sweep found no September 18 first-party publication by early Friday morning. OpenAI feed items were omitted because their direct pages could not be independently inspected through the available verifier. Coverage is selective.