Supervision is not a hire.

Supervision is not a hire.

Wednesday, October 1. Classie put Supervise into general availability. CIOs and CISOs get one view of agent activity across endpoints, browsers, and enterprise compute. Sanctioned agents and unsanctioned ones. The platform builds a runtime transcript as the workflow happens: who initiated it, what the agent did, what data it touched, what it cost. Open Policy Agent rules enforced by lightweight sensors. Alert, restrict, reroute, or stop before the action becomes an incident. Install in fifteen minutes. Monitored posture inside five days, they say. Discover shipped in April. Analyze in July. Supervise closes the sequence: find what is running, understand how it behaves, then intervene.

Two days earlier, September 29. MongoDB launched Atlas Agent Engine at its Investor Day at the Nasdaq MarketSite. Unified execution, memory, and governance for production agents. Public preview at agentengine.mongodb.com. Consumption pricing against existing Atlas commitments. Pablo Stern-Plaza named the false tradeoff enterprises keep getting: lock into one vendor's runtime and model, or stitch governance and memory yourself and watch it break every time the stack moves. Same day Oracle announced Fusion Claw, a governed agentic execution runtime for Fusion Agentic Applications. Twenty-five new Claw-powered apps, on top of a portfolio that already sits at seventy-five. Enterprise Operating Envelope: objectives, SOPs, policies, permissions, escalation boundaries. Outcome Trust Harness on each run. Outcome Receipt when it finishes. Same day Relevance AI shipped Invent 2.0, a collaborative agent that interviews you, maps the process, and builds a coordinated workforce with testing and monitoring baked in. Secure Code Warrior's deal desk now runs on a small number of Invent prompts, per their VP of FP&A.

September 30. Neutrinos launched Kamios at InsureTech Connect in Las Vegas. Governed orchestration for insurance and regulated industries. Agents on live cases: claims, underwriting, onboarding, credit. Every action governed, metered, attributable. Runs on the buyer's infrastructure, buyer's region, buyer's model choice. They put a panel on stage titled what agentic AI actually costs to run.

Last week the buyer side wired agents into core ops and discovered the procedure object still had no owner. This week is different again. Half the ships are hire-path stacks. Memory. Envelopes. Receipts. Workforce builders. The other half is Classie admitting the sprawl already happened, including agents nobody put on a purchase order.

The read everyone had

Default take: governance week. Everyone bolted policy, audit, and human-in-the-loop onto agents so CIOs can sleep. MongoDB ends the stitch-or-lock-in choice. Oracle gives you an envelope. Classie gives you a console. Relevance builds the workforce for you. Kamios meters the case.

That read is right about the nouns and wrong about the commercial act.

These are not the same instrument. One set helps you hire an agent into a stack you already pay for. The other finds agents that showed up without a hire and tries to contain them after the tokens already moved.

A transcript of the shadow is not a counterparty

Restate Classie without the press kit. Employees and developers are already running agents the company did not license. Personal AI apps. Harnesses nobody inventoried. Sessions that touch files the security team had not seen. Classie's healthcare example: fifteen minutes to install, twenty-four hours to surface known and previously unknown tools, including sensitive data going into applications the firm had neither licensed nor sanctioned. Their high-growth AI customer saw every live session in five minutes, including harnesses it had not tracked.

That is not a deployment story. That is an after-action story with a nicer UI.

MongoDB, Oracle, Relevance, and Kamios are arguing about how to put a hired agent into production with memory, envelopes, receipts, and meters. Classie is arguing that some of the agents already acting inside the firm were never hired. Different missing object. A purchase order, a principal, a named owner who meant for that agent to exist. A transcript that reconstructs what the shadow did is useful. It is not the same as having bought the work.

i keep landing on how easy it is to confuse those two. Both get filed under "governance." Both produce logs. Both claim real-time control. One starts with a decision to employ an agent. The other starts after the agent already spent.

I've watched physical markets make the same category error. You can put cameras on a warehouse full of inventory you never ordered. The footage is clean. The receiving dock still has no PO. Agents compress that failure. They do not invent it.

The machine-speed problem

Shadow software used to be a SaaS login on a personal card. Ugly. Recoverable. Someone noticed the invoice. Revoked the seat. Wrote a policy nobody read.

An unsanctioned agent that writes code, moves files, or hits production APIs does not leave that slack. By the time Discover finishes its five-day posture map, the spend is attributed and the actions are already in the transcript. Classie's attributable spend feature is honest about the order: see where the money went, then decide what to stop. Oracle's Outcome Receipt is honest about a different order: authority applied, then evidence, then result. Those are opposite sequences wearing the same word.

I wonder if the first expensive fight this quarter is not a rogue hired agent that escaped its envelope, but a clean Classie transcript of an agent nobody authorized, attached to a person who "just wanted to finish the close," with spend that already cleared. The supervision will be excellent. The hire never happened.

What the right layer looks like

What the buyer needs is thinner than another governance console. A hire object: who decided this agent exists, under which principal, against which budget, with which stop conditions, before the first tool call. Independent of whether MongoDB held the memory, Oracle held the envelope, or Classie held the sensor. Aimed at the people who have to answer for the spend and the action: finance, security, the process owner from last week's ghost-in-the-org-chart problem.

Invent's "concept to production in hours" language is the hire path at full speed. Fine when someone meant to buy the workforce. Dangerous when the same speed language gets applied to containing agents that arrived as browser extensions and personal accounts. Kamios metering per case is the hire path with a unit of account. Classie spend attribution is the containment path with a unit of account. Same unit. Different contract.

Watch who refuses to call containment a hire. Everything else this week is still a better way to log work you may not have meant to buy.

The order is always wrong

#027 was refuse on identity. Seller side. Can I see the agent before I let it in.

#028 was procedure ownership. Buyer side. Agents executing SOPs nobody versioned.

This week is the sprawl admitting itself. Hire-path platforms shipping envelopes and memory. Containment platforms shipping transcripts of agents that never got a PO. Capability and curiosity shipped first. The hire object that travels with the run, checkable when the token burns, is still the layer most of the market prices as a dashboard feature instead of a commercial act.

I've watched that sequence enough times to stop being surprised. Tools show up. Spend shows up. Governance arrives as supervision of what already ran. The purchase order shows up last, if at all, and only after someone has to explain the bill.

Close

Wednesday Classie supervised sanctioned and unsanctioned agents and built the transcript as they worked. Monday MongoDB put memory and governance in one engine. Oracle put an envelope and a receipt on complex Fusion work. Relevance built the workforce for you. Tuesday Neutrinos metered the regulated case.

Last week naming the agent was not enough if the procedure had no owner. This week supervising the agent is not enough if nobody hired it.

Watch who treats the hire as a commercial object, not a log line. The sprawl already ran. Containment is not a purchase order.

Max

Disclosure: Accords, where I'm Executive Chairman, is a portfolio position I chair.

Max Curnin has spent 11 years and 45+ investments on the physical rebuild of the economy: energy, compute, robotics, and the materials underneath. He writes The Agent Economy on the category he invests in.