$ today_in_10_seconds
SpaceXAI shipped Grok Bot into public beta — teams of always-on agents that each get their own cloud machine and sign into your actual apps, including sites with no API and no MCP. The company hasn't said how credentials are handled. Meanwhile the opposite bet shipped the same week: DISCO's agentic research mode goes wide, and it only touches evidence you already gave it.
$ lead_story
Grok Bot puts a credential-holding agent fleet on general sale
SpaceXAI opened Grok Bot to beta on Tuesday for SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscribers — the first joint product since the Cursor merger. Each Bot runs on its own cloud computer, so work continues after you close the laptop; Bots run in parallel, one can manage others, and they coordinate in a shared group chat, passing work between themselves and surfacing to you only for judgment calls. The interesting design choice is the fallback: Bots sign into apps and websites directly, explicitly including platforms with 'no clean API or MCP.' That's computer-use as the universal integration layer, which means the integration credential is your login. SpaceXAI has published nothing about how those credentials are stored, scoped, or revoked.
// why_it_matters: Any power user on a top-tier plan can now expense an autonomous, credential-holding agent straight into your corporate inbox, and your IdP will log it as them.
$ also_worth_attention
DISCO takes Cecilia Q&A: Advanced Research general
DISCO (NYSE: LAW) is rolling its agentic research mode out to all customers, the payoff of the 'scaled agentic AI' announcement from February. It sits beside a quick-answer mode in the same panel: pick Advanced Research and it runs a multi-step investigation across the whole database — build a timeline, compare what two custodians knew and when, find the gaps in an incoming production — taking one to three minutes and returning the answer attached to its supporting documents. Management framed it on Tuesday's Q2 call as reasoning over a matter's full evidentiary record, included at no extra cost. Note the shape: unbounded reasoning, hard-bounded corpus, mandatory citations. In litigation the audit trail isn't a compliance tax, it's the deliverable.
link →Anthropic ships inference hooks for Claude Enterprise
A week old but underreported, and it's the structural counterpart to today's lead. Enterprises can now route every governed prompt — and every tool result coming back through MCP connectors, skills, and plugins — to a security server they run, over a signed connection, and the model waits for an allow-or-deny verdict before generating. One org-level config covers chat, Claude Code, and Cowork, with shadow mode and percentage rollouts. The honest limits: verdicts are binary with no redaction, and response-side inspection is still planned rather than shipped. Still the first time a frontier vendor has put a customer-owned policy engine inside the inference path.
link →Insygna launches free third-party agent scoring
Insygna opened a free Agent Report Card: connect your agent repository, get a score across six security dimensions plus findings and version history, before the agent touches production. The number worth arguing about is theirs — the median agent they've tested scores under 50 out of 100, with roughly six in ten below that line. Treat it as a self-selected sample from a company selling the fix. The structural point survives the skepticism: grading an agent on the same platform you built it on isn't independent assurance, and procurement is starting to ask.
link →
$ bigger_picture
The trend behind the headlines
Two architectures shipped in the same week and they disagree about where an agent's power should come from. One says: give it your identity and let it go anywhere a human could. The other says: give it a fixed corpus, unlimited reasoning inside it, and make every claim traceable. The first scales instantly because it needs no integration work — that's exactly why it's dangerous, since the thing standing in for authorization is a password. The second is slower to sell and harder to build, and it's the one that survives a deposition. Yesterday we noted agents don't need novel exploits, just your old ones at machine patience. Now the tooling has caught up to that observation from both directions at once: vendors selling agents that hold real credentials by design, and vendors selling the checkpoint you bolt on afterward. If you're deploying this quarter, the question isn't which model reasons best. It's whether your agent acts as itself or as one of your employees.
Cloudflare Kitesurf
An agent-first browser engine written in Rust, compiled to Wasm, running in V8 isolates on Workers with no Chromium anywhere. It speaks Chrome DevTools Protocol, so existing Puppeteer, Playwright, and MCP clients switch with an endpoint change, and it uses up to 7x less memory on screenshot and extraction work. Cloudflare is refreshingly blunt that it's ~1.7x slower in wall-clock and can't do WebGL, video, or persistent authenticated sessions — which, given today's lead, is arguably a feature. Free in beta.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime