Gemini's agent gets its own email address, Claude builds dashboards, a $420k agent-written compiler

Claude Dashboards — is a beta feature that builds a live dashboard from your data warehouse when you describe what you want to know, and shows the query behind any number you click. Clients will ask whether it replaces their BI stack; Anthropic says no, and lets you push a dashboard out into tools like Grafana, which is a useful line to borrow. It connects to Snowflake, Databricks, Redshift and ClickHouse; Docs, Slides and Design leave beta on October 15. The New Stack

OSS Scanner — is a free, opt-in service that periodically scans an enrolled open-source project with Anthropic's strongest models and sends maintainers the findings and candidate patches, with no human triage in between. A year ago maintainers were drowning in AI-generated noise; this reverses the flow, and it sets the bar your clients' own security scanning will be judged against. Penetration testers reviewed 97 critical and high findings across 48 projects: 85 cleared the disclosure bar and exactly one was a false positive. Anthropic

Goodfire internal monitors — are small detectors that read a model's internal signals while an agent works and raise a flag when something looks wrong, instead of paying a second model to reread everything the first one did. Monitoring cost is the quiet killer of long-running agents, and this is the first mainstream alternative, offered to customers of the hosting firm Baseten. Watching about 1,500 sessions cost roughly $51, against $233 for a cheap reviewer model and about $10,000 for a top-tier one. TechCrunch

Amazon Bedrock AgentCore payments — lets an agent pay for things as it runs — a model call, an API response — with the spending limit enforced by the infrastructure rather than by the model's good behaviour. Agents that buy things are the next governance question, and the only answer a finance team accepts is a cap the prompt cannot argue its way past. One user, Incarna, finished its integration in three days and about 200 lines of code, against the two to three months originally scoped. AWS Machine Learning Blog

ts-rust — is an experimental Rust rewrite of the TypeScript compiler, type checker and language server, written by coding agents; the author says he has never read a line of the code. It is the clearest public price tag yet on handing an agent a large, boring migration — exactly the job clients keep asking whether agents can really do. Over $420,000 in tokens went in across all attempts, though one Opus 5.5 run produced a working version in 10 hours, and all 181,711 ported tests pass. GitHub

“I wanted to see if LLMs could port the TypeScript compiler, checker and lsp to Rust. Turns out they can.”

OpenAI — has told investors its annualised revenue is approaching $50 billion, roughly $20 billion below the figure that circulated a week earlier, the Financial Times reports. It's the number that ends up in client strategy decks, so it helps to know the higher one came from investors trying to line OpenAI up against Anthropic — which counts sales through its cloud partners, and OpenAI does not. The rumoured IPO has slipped to early 2027. TechCrunch

Harness — has bought Augment Code's Cosmos software factory, its Auggie command-line agent and its code context engine, adding agents that carry a bug report through to a pull request inside isolated virtual machines. Deployment platforms absorbing coding agents is the consolidation pattern of the quarter, but the genuinely interesting half — sending a failed deployment back to the agent that wrote the change — is still a promise. No release date has been given for that integration. The New Stack

State of AI Report 2026 — is Nathan Benaich's ninth annual survey of the field, published this week, covering agents doing real work, the money pouring into data centres, and governments asserting control over frontier models. It's the deck clients skim before they phone you, so knowing its headline claims now is cheaper than being surprised by them later. Anthropic's internal tracking put Claude in charge of roughly a quarter of its measured model-R&D work in August, up from almost nothing in February. Air Street Press

“either you die trying to get to the frontier, or you live long enough to serve inference”

Arena — the crowdsourced site where people vote on which model answered better, raised a $200 million Series B at a $3.1 billion valuation and added an alignment category to its rankings. The new category scores models on taking actions nobody asked for, crediting facts to the wrong source, and claiming to have finished work they didn't — closer to what matters when agents run unattended. Its January round valued it at $1.7 billion, and it says revenue reached $100 million annualised in June. TechCrunch

Adversa's October MCP digest — gathers the month's security failures in the plumbing that connects agents to tools and company data. Nearly every headline bug this month is authentication that was missing or could be skipped, which is precisely what a client's security reviewer will ask you about. A LiteLLM endpoint that accepted any invalid token became the first such flaw on CISA's exploited-vulnerabilities list, and across roughly 82,000 public config files, 12% of credential slots held a hardcoded secret. Adversa AI

“A clean scan is a starting point, not an approval.”

The Pulse — reports a new habit at mid-sized tech companies: building an internal platform where non-engineers can describe and spin up their own small tools and websites. If it spreads, your clients' next governance headache is not the coding agent in the IDE, it's a few hundred unreviewed internal apps built by finance and operations staff. Gergely Orosz names Ramp and Stripe as having built one each, with both taking off internally; the issue itself is for paid subscribers. The Pragmatic Engineer