Capital chases the agent interface —
Hark, Exa, and agents in your apps.

Brett Adcock — the founder behind Figure (humanoid robots) and Archer (eVTOL) — raised over $700M in a Series A at a $6B post-money valuation for Hark, an AI lab building both foundation models and custom hardware for a "universal" personal AI interface: one agent that sits between you and every digital system. The round was led by Parkway Venture Capital with Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures and ARK among the backers — a chip-and-cloud roster that signals what they think the interface layer is worth owning. Hark says its first models ship later this summer. Keep the frame honest: this is a Series A, so the valuation prices the thesis, not shipped revenue. It matters precisely because that thesis — the agent, not the app, becomes the surface — is the one every incumbent is now defending against.

Exa Labs builds a search engine designed for agents rather than people — structured, grounded web access over an API, with its fastest tier (Exa Instant) returning results in under 180ms. The $250M Series C, led by Andreessen Horowitz, more than triples the valuation from its raise last fall ($85M at $700M). The money goes to training its next-generation retrieval models and to infrastructure built to handle "hundreds of thousands of searches per second." Read it alongside Parallel's $2B round three weeks earlier: web access is becoming a paid, agent-specific utility layer, and Google and Bing no longer own it by default.

London's Primer raised €86.2M (Series C, led by Sofina and Peak XV) to push its payment-orchestration platform toward what it calls autonomous AI payments — agents that don't just decide but transact, across 30+ countries behind a single routing and reconciliation layer. Total raised is now €146.6M. Stack it with Hark and Exa and the week's funding maps neatly onto the three things an autonomous agent actually needs: a way in (interface), a way to read the world (search), and a way to pay (rails). The agent economy's plumbing is getting funded well ahead of the agents being reliable — which is either prescient or 1999, depending on your priors.

Cursor shipped Composer 2.5, its in-house agentic coding model tuned for long-horizon, multi-step tasks. The headline is economics: it posts 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1 — matching Claude Opus 4.7 and GPT-5.5 on those benchmarks at roughly a tenth of the per-token cost, with a faster default tier and double usage in the first week. The strategic read: the frontier labs no longer hold a monopoly on "good enough for an agent loop," and the tools that own the editor can now afford to run their own model on every keystroke.

OpenAI added a first-class sandbox for Codex on Windows: dedicated setup and runner binaries, firewall-backed network blocking, and tighter file-write controls so the agent can run shell commands and edit code without unsupervised reach into the host. It's the same containment pattern Anthropic shipped for Managed Agents and Google for Antigravity last week — every vendor is converging on "the agent runs, but inside a box your security team drew." The unglamorous sandbox layer is quietly becoming the real enterprise battleground.

Figma began rolling out its native AI agent in beta: natural-language prompts that generate and remix designs, automate rote work, and respect your design system out of the box, with multiple agents runnable in parallel. The notable part is the plumbing — Figma wired in both Anthropic's Claude Code and OpenAI's Codex under the hood, a rare two-lab bet that says the design surface wants the best coding agent regardless of vendor. It landed alongside Q1 revenue of $333.4M, up 46% year-on-year — evidence that "add an agent" is, for now, accretive rather than cannibalizing. During beta the agent won't consume credits.

A local-first store (Python, MCP-compatible) that draws a distinction the memory-layer crowd has been circling: it remembers who you are — preferences, conventions, working style — not just what you did, and exposes it to Claude Code, Codex and Cursor through one MCP server. Same local-file-not-managed- service philosophy as dejavu last week, aimed one rung up the stack at identity rather than episodic recall. Early (~50 stars) but a clean articulation of how agent "memory" is splitting into sub-problems.

A privacy-first macOS menu-bar tracker that reads your local session logs to show Claude Code and Codex usage — no provider API calls, no telemetry, no account. As per-token agent spend turns into a real line item, the tooling to watch it without shipping your data anywhere is exactly the small, local utility that keeps topping the trending lists (~90 stars in days). It lives in the same local-first observability lane as weft last edition.

"Personal Overnight Helper Agent" runs while you sleep and serves a morning brief before your alarm — a tidy instance of the unattended, long-horizon pattern everyone keeps gesturing at, packaged as something a single developer can run on Claude Code tonight. It is, fittingly, a hand-built cousin of what this very brief does. Worth noting the genre: the autonomous-agent dream is arriving first as small personal cron-with-judgment tools, not enterprise platforms.

An offensive-security toolkit packaged as Claude Code skills — red-team workflows, exploit development, Active Directory attack chains, EDR-bypass and mobile-pentest playbooks (~130 stars in days). For authorized testing it's a marker of where the skills format is heading: not just productivity helpers but deep, domain-specific operator toolkits. The flip side is the obvious dual-use question — and it'll be a useful test of how the skill ecosystem handles capability that cuts both ways.

A skill collection that points an agent at distributed-systems testing: fault injection, chaos-engineering scenarios, and the failure-mode reasoning that usually lives in a senior SRE's head (~170 stars in a few days). It's the most credible example this week of skills moving past "write my commit message" into genuinely hard engineering domains — and a hint that the high-value skills will be the ones encoding scarce expertise, not common tasks.