$ today_in_10_seconds
Perplexity and Nvidia shipped a full agent stack that runs on hardware you already own — model, harness, sandbox, connectors, and a credit counter frozen at zero. It's the clearest bet yet that the thing constraining agents isn't intelligence, it's the meter. Meanwhile, the boring plumbing underneath agents — protocol governance, payments, write permissions — all moved forward in the same week.
$ lead_story
Perplexity and Nvidia ship a fully local agent with no token bill
Perplexity launched Portable Computer today, a local version of its agentic Computer platform that runs entirely on Nvidia's DGX Spark or Linux machines with RTX GPUs — model, agent harness, inference engine, tools, connectors, and security sandbox bundled into one app. Work done on-device consumes no billing credits, every task starts local by default, and the system asks permission before escalating any individual step to a frontier model in the cloud. The interesting part isn't the packaging — it's the accompanying research argument that small local models need a harness co-designed with them, because general-purpose harnesses assume a frontier model that can absorb huge contexts and navigate sprawling tool surfaces. So Perplexity built a deliberately minimal one: a short system prompt, a small core toolset, capabilities that load and unload as on-demand skills, popular MCP connectors like Gmail and GitHub rewritten as compact CLI tools, and always-on OS sandboxing that disables the harness entirely if the sandbox is unavailable. The economics are the real headline. On Terminal Bench 2.1, the local Qwen model scored 59.6% at essentially zero marginal cost; letting it escalate to a Claude Opus 5 advisor pushed it to 73.0% at roughly $0.415 per task, versus 82.4% at $0.65 running the frontier model alone. Two-thirds of the price, three-fifths of the gap closed — and you choose per task. Caveats: the headline benchmark is Perplexity's own, it's Linux-only until Windows lands in September, the 24GB VRAM floor rules out most consumer PCs, and Apple silicon isn't on the roadmap.
// why_it_matters: When inference is free, "let the agent verify its own work three more times" stops being a budget decision.
$ also_worth_attention
AWS makes agent payments generally available
AgentCore payments went GA, letting agents autonomously discover and pay for APIs, MCP servers, and content via Coinbase and Stripe Privy stablecoin wallets — with delegation from the end user, spending limits enforced at the infrastructure layer, and credentials the agent never sees. It now supports both x402 and the Machine Payment Protocol, plus an "upto" scheme for pay-per-inference pricing. Read it alongside the lead story: one launch removes the cost of thinking, the other builds rails for agents to spend money on everything else.
link →A2A moves in with MCP under one foundation
Google's Agent2Agent protocol became a hosted project of the Agentic AI Foundation, putting it in the same home as MCP — A2A handling communication between independent agents, MCP handling connections to tools and data. AAIF has grown from fewer than 40 members at its December 2025 launch to more than 250, including Google, Microsoft, Amazon, Anthropic, OpenAI, Bloomberg, Shopify, and Block. The layer split is now settled enough to build on — which matters more for procurement than for engineering.
link →Cloudflare starts gating what agents are allowed to change
WriteGuard entered private beta for Cloudflare MCP server portals, built internally before Cloudflare expanded write access across its own MCP servers — and it deliberately avoids standalone agent accounts, so an agent inherits exactly the permissions of the employee behind it. Each tool gets a risk tier: reads pass through unchanged, allowed writes get agent attribution and an audit event, critical actions are blocked before the handler runs. The framing is a permissions problem, not a threat-actor one — human privileges were bounded by judgment and human speed, and agents are bounded by neither.
link →
$ bigger_picture
The trend behind the headlines
Four announcements, one theme: nobody serious is arguing about model choice anymore. They're arguing about where the agent runs, what it's permitted to modify, and who pays for each step. Perplexity's answer is that metered inference silently caps how long an agent is allowed to think, so move it on-device. AWS's answer is that agents will spend money, so give them a wallet with a ceiling. Cloudflare's is that agents will change state, so put a gate in front of the write. These are all the same admission from different angles — that the agent harness, not the model, is where the engineering and the risk actually live. Local execution kills the token bill but doesn't kill the audit requirement, which is why the boring control-plane work will outlast the benchmark chart.
Terminal-Bench 2.1
The benchmark behind today's most useful number. It scores agents on real terminal tasks, and because it's harness-sensitive rather than purely model-sensitive, it's one of the few public leaderboards that tells you something about your scaffolding instead of just your model provider. Worth running your own stack against before you accept anyone's escalation-cost math.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime