The Lyceum: Agentic AI Weekly — Mar 17, 2026

This was the week AI agents stopped being something companies experiment with and started being something they buy in bulk. A 30,000-person staffing giant signed an unlimited license to deploy agents across every brand and country it operates in, Nvidia used GTC to unveil the plumbing it hopes every enterprise agent will run on, and Mistral collapsed four separate AI models into one — open-source, Apache-licensed — that can reason, see images, and write code without switching brains mid-task. The question has shifted: it's no longer "can agents do useful work?" but "who controls the infrastructure they run on, and what happens when they break?"

Here's what it looks like when a company stops piloting and starts committing. The Adecco Group — one of the world's largest staffing firms, placing hundreds of thousands of workers annually — announced on March 12 that it purchased unlimited global access to Salesforce's Agentforce platform through 2027. Not a hundred seats. Not a regional trial. Unlimited.

The agents will handle the high-volume, repetitive glue work of recruiting: screening candidates, scheduling interviews, chasing documents, keeping records current. Salesforce's case study frames this as augmentation — freeing human recruiters to spend time on the parts of the job that require judgment and empathy.

What makes this genuinely interesting isn't the technology; it's the contract structure. Unlimited seats mean Adecco can deploy agents everywhere simultaneously — across brands, countries, and functions — which will stress-test Agentforce's governance and data-access controls at a scale no pilot ever could. Expect this deal to become the reference case other large employers wave at their boards, unions, and regulators when arguing their own agent rollouts deserve the green light.

Nvidia doesn't just want to sell you the chips your agents run on. It wants to orchestrate the agents themselves. At GTC in San Jose this week, Nvidia formally unveiled NemoClaw, an open-source platform for deploying persistent AI agents — bots that watch data streams, trigger actions, and coordinate tools over hours or days, not just single conversations.

The platform is pitched as hardware-agnostic (you don't need Nvidia GPUs, wink wink), and the partner list is serious: Salesforce, Adobe, SAP, Cisco, CrowdStrike, and Google have all committed to building on the stack, according to VentureBeat. NemoClaw bundles a secure agent runtime with sandboxing, least-privilege access controls, and a privacy router — features that Futurum Group's analysis calls the first serious attempt at production-grade agent security from a major infrastructure vendor.

Nvidia also announced the Nemotron Coalition, a consortium including Mistral, LangChain, Perplexity, and Cursor that will co-develop open foundation models on Nvidia's compute. The strategic play is the same one that worked with CUDA: give away the orchestration layer, make the hardware the natural — and eventually indispensable — place to run it. If NemoClaw gains the kind of traction Kubernetes did for containers, Nvidia will have locked in a layer of the stack far more durable than any single GPU generation.

For the past two years, if you wanted to build an AI agent, you needed a shopping list: one model for quick chat, another for deep reasoning, a third for understanding images, a fourth for writing code. Mistral just tore up that list.

Mistral Small 4 is a 119-billion-parameter model — released under Apache 2.0, meaning anyone can download, modify, and deploy it for free — that unifies instruction-following, step-by-step reasoning, image understanding, and code generation in a single deployment. The clever trick: a reasoning_effort parameter developers can adjust per request, as Simon Willison noted, so a simple "what's the weather?" query runs fast and cheap while a complex debugging task gets the model's full analytical horsepower. No model-switching required.

The efficiency numbers matter for anyone paying agent compute bills. Mistral reports a 40% reduction in completion time and 3x more requests per second compared to its predecessor in Mistral's vendor benchmarks. Those are vendor benchmarks — salt accordingly — but the community response on Hacker News and Reddit has been genuinely enthusiastic. Running one model instead of four is an operational simplification that compounds across every deployment. If this "unified model" pattern spreads to other labs, the infrastructure decisions facing every team building agents just got meaningfully simpler.

Here's a very 2026 sentence: your weekend walks catching Pikachu may have been teaching delivery robots how to navigate your sidewalk.

MIT Technology Review reported that Niantic, the maker of Pokémon Go, used over 30 billion images captured by players to train a visual positioning system now powering small delivery robots. As players roamed cities pointing phones at landmarks to catch Pokémon, Niantic collected dense, ground-level imagery of sidewalks, storefronts, and intersections — the exact view a delivery bot needs to navigate where GPS fails. That dataset captures variations in weather, lighting, angles, and seasons that Popular Science notes staged robotics datasets simply can't manufacture.

The first customer: Coco Robotics, which operates about 1,000 small sidewalk delivery robots across Los Angeles, Chicago, Miami, and other cities, with over 500,000 deliveries completed, per TalkEsport. Because the dataset contains many images of the same locations from thousands of different users, it provides the robustness robotics companies need for reliable navigation.

For the agentic AI world, this is a reminder that the most powerful autonomous systems often ride on data exhaust from consumer apps — and that debates over consent, compensation, and who actually owns that data are about to get much louder as more robots roll down our streets.

Identity management for human workers is a solved problem. For autonomous AI workers? Not even close. Okta — the company that manages login and access for millions of employees at thousands of organizations — published this week a security blueprint and preview product that treats each AI agent as a non-human identity with its own role-based permissions and an emergency kill switch.

The logic is straightforward: as companies deploy fleets of agents that access apps, databases, and APIs, someone needs a central control plane to see which agents are acting, what permissions they hold, and to revoke access instantly if behavior goes sideways. Okta is positioning itself as that control plane.

This matters because it signals the birth of a new product category — agent identity and access management — and sets expectations for how security tools will integrate with agent runtimes. A viral Reddit thread claiming 64% of billion-dollar enterprises lost over $1 million to AI failures last year (treat that number as community signal, not verified data) suggests the demand side is already anxious. When Okta moves, CISOs notice — and procurement requirements tend to follow.

A staffing company buying unlimited robot recruiters like a Costco membership. A GPU maker giving away the agent operating system so you'll keep buying the hardware. A decade of Pokémon walks quietly becoming the world's most valuable sidewalk map.

Somewhere, a Pikachu is filing a class-action lawsuit for unpaid data labor — and Okta is asking it to verify its identity before it can retain counsel.

Until next week, watch the agents. Someone should.

If you know someone who'd enjoy this, send it their way.

The FTC's new enforcement stance means "unfair" AI practices are now prosecutable — U.S. AI policy just moved from white papers to real legal risk. Read → FTC Draws a Line: "Unfair" AI Is Now an Enforcement Target Under Section 5