← /editions
#81

The token meter just hit zero

// Perplexity and Nvidia ship a full agent stack that runs on your own GPU with zero cloud token cost, while Binance opens trading rails to agents, A2A joins MCP under one foundation, and Google wires agents into law firms. Capability is cheap; provenance is expensive.

$ today_in_10_seconds

Perplexity and Nvidia shipped a full agent stack that runs on hardware you already own, with the cloud credit counter parked at zero. Meanwhile Binance handed agents the keys to real money, Google shipped agents into law firms, and A2A moved in with MCP under one governance roof. Four stories, one theme: the interesting fights are now about cost, authority, and plumbing, not capability.

$ lead_story

Perplexity's Portable Computer moves the whole agent stack on-device

Perplexity launched Portable Computer today, a version of its agentic Computer platform that runs entirely on hardware users already own, starting with Nvidia's DGX Spark and Linux machines with Nvidia RTX GPUs. This isn't a local chatbot with a cloud fallback: the orchestration layer, planner, tool selector, scheduler, persistent task queue and local search index all execute on the device, and every task begins locally, with the system asking permission before shipping an individual step to a frontier model in the cloud. The demo money shot was the credits counter, which Perplexity's VP of engineering described as parked at zero because all of it ran on the box. Read the fine print before you re-forecast your inference budget: you get a 27B Qwen model or Perplexity's own PPLX 27B, with Nemotron 3.5 Lightning to follow and bring-your-own-model supported, but it needs an Nvidia GPU with at least 24GB of VRAM, ships Linux-first, gets Windows in September, and has no announced macOS support. A 27B model is not a frontier model, which means escalation-to-cloud will fire far more often in your workflows than in a scripted demo. Zero token cost is really capex-for-opex: you pre-paid for the tokens when you bought the GPU.

// why_it_matters: Agent unit economics just became a hardware procurement decision, and data residency finally has a product to point at instead of a policy document.

read more →

$ also_worth_attention

  • Binance opens its trading rails to agents

    Binance launched Agent OS on August 20, a developer platform letting AI agents access market data, monitor accounts, execute trades, and touch payments and on-chain wallets, with activity isolated via user-configured subaccounts and permission limits. Agents get dedicated subaccounts with withdrawals blocked by default and daily caps of $50,000 for swaps, $100,000 for DeFi and $20 for x402 payments. The honest disclosure is the important one: Binance can monitor trades placed through Agent OS but cannot see an agent's external information sources, interpretation or decision-making, which happen inside the user's own AI application. So the exchange can prove what an agent did and never why, which is exactly the gap every financial-services compliance team is going to ask you about.

    link →
  • A2A moves in with MCP

    Google's Agent2Agent protocol is becoming a hosted project of the Agentic AI Foundation, putting it in the same home as MCP, which covers agent-to-tool connections while A2A handles communication between independent agents. AAIF says it 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 protocol wars are effectively over, which is good news and also a signal: differentiation now moves up-stack to identity, authorization and audit, the layers neither protocol actually solves for you.

    link →
  • Google puts agents inside law firms

    Gemini Enterprise for Legal shipped today with purpose-built skills for brief drafting, citation verification, contract lifecycle management, regulatory horizon scanning and DSAR fulfilment, plus MCP connectors into Harvey, Thomson Reuters, iManage, NetDocuments, RelativityOne, Everlaw, Docusign and CourtListener. Weil Gotshal, Cleary Gottlieb, Freshfields and Williams and Connolly are among the firms collaborating with Google, and Reuters describes the bundled agents as handling specialized legal and administrative functions without significant human oversight. Note what the product actually is: connectors and skills, not a model. That's the vertical agent playbook, and without significant human oversight is a phrase a general counsel is going to want rewritten.

    link →

$ bigger_picture

The trend behind the headlines

Nothing shipped today because a model got smarter. Perplexity moved where the compute runs, Binance defined who authorizes an action and up to what limit, AAIF settled which wire the request travels on, and Google wired agents into the systems of record lawyers already trust. That is four different bets on the unglamorous middle of the stack, and it's a reasonable read on where the margin now lives. It also exposes the shared weak spot: Binance admits it can't see agent reasoning, Google's legal agents run with light oversight, and a local 27B model will silently escalate to someone else's cloud. Capability is cheap, provenance is expensive, and nobody has shipped the audit trail yet.

$ tool_of_the_day

NVIDIA NeMo Agent Toolkit

If today convinced you to model your agent economics properly, this is the instrumentation layer. It's an open-source, framework-agnostic library that profiles entire workflows from the agent level down to individual tokens to find bottlenecks and analyze token efficiency. It works alongside LangChain, LlamaIndex, CrewAI, Semantic Kernel and Google ADK, with observability hooks for Phoenix, Weave, Langfuse and OpenTelemetry, plus full MCP client and server support. Profile before you decide whether local inference actually saves you anything.

explore NVIDIA NeMo Agent Toolkit →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime