Enterprise agents take the search bar —
and Modal takes $355M.

Marc Benioff announced Coworker on X late Thursday, framing it as "your new AI teammate inside every search bar." The beta plugs an Agentforce agent into Salesforce Global Search across Enterprise, Unlimited and Agentforce 1 editions; users ask in natural language, the agent reads CRM context — opportunities, cases, customer history — and takes action without the tab-hunting and copy-paste that defined the old workflow. It's available today for every Agentforce customer. The strategic read is the surface, not the model: Salesforce isn't selling a chat window any more, it's quietly making the search bar agentic for every CRM seat it already owns. That's the move every SaaS incumbent is going to copy this quarter.

Computer use in Copilot Studio graduated from preview to GA and rolled out to every commercial Power Platform geography, with the rest of the month spent landing in tenants. The agents do not drive selectors — they read the screen and reason about what to click next, which is what lets them survive layout changes, branching flows and legacy UIs that don't expose an API. Microsoft made governance the loud part of the message: audit trails, human approval checkpoints, secure auth and observability are wired in by default, run through the Power Platform admin centre. The interesting consequence: the cheapest way to "give the agent a screen" inside a Fortune 500 just stopped being a research project and became a paid-tier toggle on the platform IT already owns.

Camunda announced ProcessOS at CamundaCon in Amsterdam to 1,200 enterprise customers: an intelligence layer that sits inside the Camunda Hub and runs four agents across the process lifecycle — discover how the org actually runs today, re-engineer toward target outcomes, build and deploy, then continuously improve once live. Closed beta opens to selected enterprises now; ProcessOS runs natively on AWS with deep hooks into Bedrock and Bedrock AgentCore for memory, identity and the gateway. The line worth quoting from the launch: "AI-first process transformation must work backwards from outcomes, not forward from current reality" — a pointed shot at every "AI bolted onto our legacy BPMN" pitch deck out there. BPM was the category that looked least promising for the agent thesis; turning each lifecycle stage into a specialist agent is the cleanest answer we've seen.

Modal closed $355M led by General Catalyst and Redpoint, with Menlo, Bain Capital Ventures and Accel joining, at a $4.65B post — up from roughly $1.1B in September. CEO Erik Bernhardsson pegged annualised revenue at ~$300M, up from $60M in the same window. The product Modal is actually selling is the layer that everyone else's announcements ran on this month: serverless sandboxes for tool execution, low-latency inference, parallel RL loops. It's the same Modal that Anthropic listed as a partner sandbox provider for Managed Agents two weeks back. New money goes into low-latency inference primitives, the RL training stack, and parallel sandbox capacity — i.e. the things that make the agents above this section actually run. The interesting valuation question is no longer "what's the model worth"; it's "what's the runtime under the model worth, and is it more durable."

Index led an $78M Series C with a16z, Avra and the existing book following on. The headline isn't the cheque; it's the customer line. Moment now powers investment-management workflows at firms running more than $10T in client assets — Edward Jones, LPL Financial, Hightower — up from $3T six months ago and $300B eighteen months ago. The pitch is "AI operating system" for portfolio managers, advisers and traders: the safety rails (compliance, supervision, audit) shipped before the agent surfaces did, which is how it got past the regulator-facing buyers in the first place. Vertical agent stacks aren't a niche any more — they're the way regulated industries finally let an LLM near a trade ticket.

A four-year alliance combining grant funding, Claude usage credits and engineering time, aimed at three buckets: global health (the largest), K-12 tutoring and adult workforce guidance in the US plus literacy and numeracy programs in sub-Saharan Africa and India, and agricultural productivity tooling for the ~2B people who farm smallholdings. The pact includes Claude fine-tuning on local crops and benchmark sets the Foundation will help build. Read it next to the $900B round and the picture is consistent — Anthropic is building the policy and distribution coverage that lets the frontier-lab category survive in the form of a public utility, not just a Big-Tech vendor. The Gates Foundation almost never co-brands with a single AI lab; that it did here is the news.

Perplexity, of all companies, published a Go tool that does one boring thing well: read-only inventory of the packages, extensions and dev-tool metadata sitting on a developer's machine, cross-referenced against known supply-chain compromise lists. ~1.2k stars in four days. Notably not trying to be an agent — it's the thing you run on the boxes the agents run on, after a quarter in which npm, browser-extension and IDE-extension breaches kept landing in coding-agent context windows. The signal isn't the tool; it's that the lab whose product is web search felt the need to ship it. The supply-chain-of-the-agent surface is now actively defended.

Paradigm — usually known for Reth and the rest of the Ethereum tooling stack — open-sourced a Python runtime called Centaur with a deliberately uncommercial pitch: agents that are multiplayer (multiple humans in the same agent session), self-hosted, and secure by construction. ~430 stars, two companion repos for an Argo-CD-style GitOps deployment template. The interesting contrast is with the SaaS lane above: Paradigm is wagering that the orgs running serious agent infrastructure will want the same self-hosted-and-auditable shape they already insisted on for their Ethereum nodes. Worth tracking for whether the on-prem agent-runtime category gets a credible open base before the hyperscalers fully colonise it.

Simone Margaritelli — long-time offensive-security author behind bettercap and the original code-audit — shipped a successor agent that runs an explicit 8-stage pipeline over a target codebase, from reconnaissance to triage to write-up. Python, MIT, ~440 stars in five days. It's deliberately not a single-prompt "find me a bug" loop; the staging is the product, and it's the cleanest open answer yet to the white-paper "agent finds CVEs" demos the big labs have been running. Pair with bumblebee above and the picture is the same picture: the offensive and defensive tooling around the agent is now both shipping in the open.

Lightspeed-led seed-extension into Ocean, whose product "Ray" investigates suspicious emails end-to-end — fetching headers, sandboxing attachments, scoring intent against tenant-specific policy, escalating only when human review is actually needed. Already scanning over 1B messages/month for Fortune 500 customers including KAYAK, Kingston and Headspace, with Wiz and Armis founders angeling in. The category to watch isn't "AI security" generally; it's agentic-vs-agentic — phishing is increasingly LLM-authored, and the only economical defence is something that triages at machine speed against the same loop. Ocean's the cleanest first instance.

Four versions in four days, with two changes worth the upgrade. /usage now reports a per-category breakdown of what's actually driving your limits — skills vs. subagents vs. plugins vs. each MCP server — which is the first time the bill is itemised at the surface where you spend it. /simplify was renamed to /code-review and now takes --effort high plus --comment, which posts findings as inline GitHub PR comments — the review-as-skill flow we covered last week, now first-class. Pinned background sessions (Ctrl+T) survive idle, restart in place for updates, and only get shed under memory pressure after non-pinned ones. Markdown now renders GFM task-list checkboxes; enterprises get allowAllClaudeAiMcps for cloud connectors. .148 hot-fixed a Bash exit-127 regression from .147 — read the changelog before you upgrade past .146.

Browser Use shipped "Box": a small Python project that pins a Claude Code agent to a long-running box you own, wires a real Playwright browser through Browser Use Cloud, and lets you hand it tasks over Telegram. ~340 stars, MIT. The interesting bit is the deployment model — not a hosted SaaS, not a local CLI, but a personal cron-tier agent that watches a real browser session 24/7 and reports back to your phone. It rhymes with the OpenAI Codex "remote-after-lock" pattern we covered last week from a different angle: the laptop is no longer where the agent lives.

A Go project (~600 stars) staking out a niche we haven't seen elsewhere: not "agent-friendly git," but a parallel VCS whose primitives are agent runs — what context was loaded, which tools were called, which files changed as a result, which approvals fired. The bet is that "the commit" is no longer the unit of work worth tracking; the agent run is, and the right granularity for blame, rollback and review is the trajectory, not the diff. Early, opinionated, and pointed at a real gap — most teams shipping agents are still reconstructing what happened from terminal scrollback. If even half the primitives stick, this is the kind of plumbing that eats Git's space for agent work the same way Git ate SVN's for human work.