Skip to content

Knowledge Strategy

Multi-agent systems need a shared world, not just shared messages

The proposal window for the Defense Advanced Research Projects Agency's (DARPA) Decentralized Artificial Intelligence through Controlled Emergence program closed yesterday. DICE asks a difficult systems question: can heterogeneous artificial intelligence agents coordinate through peer-to-peer interaction, adapt when individual agents fail or become compromised, and still remain aligned with commander's intent over long missions?

Two days earlier, researchers released a preprint describing an “ontology as a kernel” for language-model agents. The proposed system makes domain concepts, relationships, evidence, and permissible reasoning operations explicit instead of leaving all of them implicit in prompts and unstructured context.

Those developments come from different research communities, and neither proves the other's architecture. Together, they expose the same design boundary: a collection of agents does not become a system merely because the agents can exchange messages. It becomes a system when they can coordinate around a shared, inspectable, and governed model of the world and the work.

That is a semantic-systems problem.

A follower is not an audience

Earlier today I posted something deliberately blunt on LinkedIn. I said the feed increasingly felt less like a professional network and more like Facebook: crowded with clickbait, personal-brand theater, and content optimized to provoke a reaction rather than help someone find work, exchange expertise, or make a useful professional connection.

Three hours later, the post showed 42 impressions. My profile has roughly 16,500 followers. Two people had reacted. Above the post, LinkedIn displayed an invitation to Boost it.

I took a screenshot because the juxtaposition was almost too perfect. A platform had offered me a number suggesting extraordinarily weak organic distribution and, in the same interface, offered to sell me more visibility. Then I wrote a second post criticizing that result. That complaint reached about 250 impressions in its first hour—roughly six times the number shown on the original post.

The easiest conclusion is also the most emotionally satisfying: LinkedIn suppressed my post so it could charge me to reach the audience I had already built.

The screenshot does not prove that. It reveals something more defensible, and in some ways more consequential: a follower on a ranked platform is not an addressable audience. It is a relationship that the platform may or may not activate, according to objectives the user cannot see.

That distinction matters on any social network. It matters more on a platform that mediates employment, expertise, reputation, recruiting, sales, and access to professional opportunity.

The first deliverable from generative coding should be understanding

The most dangerous sentence in a modernization program may be, “We know what this system does.”

Usually, someone knows what the system is supposed to do. Operators know the screens and workarounds. A few engineers know where the brittle integrations live. Program managers know the contracts and milestones. Cybersecurity teams know some of the exposed surfaces. The source code knows all of it at once—but in a form no single person can hold in mind.

That is why the emerging market for generative coding matters to government for a reason deeper than writing software faster. Its first serious public-sector use may be helping an organization recover a working model of the technology estate it already owns.

The CMMC suspension is an organizational knowledge test

Cybersecurity certification was never going to be just another audit for the companies that build and support U.S. defense systems. Preparing for it changed budgets, architectures, hiring plans, subcontractor relationships, and the everyday work connecting cyber teams with engineers, contracts staff, and program leaders.

Then, on July 13, 2026, the Department of Defense (DoD) suspended the planned transition to Phase II of the Cybersecurity Maturity Model Certification (CMMC) program. The Department also halted later implementation milestones and opened a 60-day review of the program.

The announcement brought meaningful relief from the coming expansion of third-party assessments. It did not, however, abolish CMMC or erase the contractual duty to protect defense information. Phase I continues, and the underlying safeguarding and reporting requirements remain in force.

That leaves defense organizations with a harder question than whether to keep preparing for an assessment: Which parts of the CMMC effort were merely certification overhead, and which parts became necessary organizational capability?

The distinction matters because a compliance program leaves behind more than policies and evidence. It also changes who talks to whom, how work moves between teams, where decisions are recorded, and whether an organization can accurately explain the security of its systems. Those capabilities are slow to build and surprisingly easy to lose.

Provenance has to survive the workflow

OpenAI has announced an expanded approach to content provenance, combining Content Credentials, SynthID watermarking, and an early verification tool. The work is meant to help people understand whether media came from an artificial intelligence system and how it may have been created or edited.

That is valuable context. Its usefulness depends on whether the context remains attached as content moves through the ordinary machinery of work.

Monitoring deployed AI is a knowledge practice

The National Institute of Standards and Technology (NIST) has published a report on the challenges of monitoring deployed artificial intelligence systems. It addresses a growing operational reality: predeployment testing cannot anticipate every combination of user, data, environment, and system change.

Monitoring is the bridge between what a team expected and what the deployed system is actually doing. Building that bridge requires more than a dashboard.

Stateful agents make memory a governance problem

OpenAI and Amazon have announced a strategic partnership that includes plans to co-develop a stateful runtime environment for artificial intelligence agents. The phrase “stateful” deserves attention. An agent that can preserve context across steps and sessions may be more useful than one that repeatedly starts from zero.

It may also accumulate assumptions, permissions, and mistakes that no one intended to become durable.

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed