Learn AI deeply. Question it honestly.
Build it for people.
What's materially changed?
Sept. 15, 2026 · NIST and CISA
Sept. 15: NIST and CISA finalized guidance on protecting tokens, the digital passes that let people, services and AI agents into systems without a password each time. The report, NIST IR 8587, shows how to keep tokens from being forged, stolen or misused, and it explicitly covers AI agents.
Sept. 29: An executive order replaced "artificial intelligence" with "super intelligence" in federal language. NIST renamed its AI center CAISSI but left its frameworks and guidance unchanged.
Every AI agent that reaches a company system carries a token. Stolen tokens caused last issue's breach of more than 700 Salesforce environments. The guidance is voluntary for private companies, but expect cloud vendors and auditors to measure against it. The meta-harness in this issue's Deep End raises the stakes: it often brokers credentials for every agent beneath it.
Washington now treats AI agents as identities, not app features. IR 8587 applies the same token rules to agents as to people and services.
The guidance protects the token, not what each agent may do. NIST says that guidance is coming. What NIST calls signed intent, proof that a person authorized an agent's action, also needs more standards work. Most companies aren't ready either: 78% have no documented policy for creating or removing AI identities.
Read more: NIST's announcement and the Cloud Security Alliance on what it means for AI agents.
"Do we know every AI agent we run and what each one can reach?"
Real term, or did we make it up?
The three games in this issue are just for fun. Guess wildly and see how you do.
This month: Meta-Harnesses
Last issue covered how MCP standardized the way agents reach systems. This issue goes one layer up, to how agents are defined and governed apart from the harness that runs them. As always, the goal is enough context to know when this is useful, when to invest and when to push back on the hype.
For two years, the question in AI was which model is most capable. September made that question look dated. At least nine labs shipped new models in a single month (see Tool Directory's September roundup). And the price tags stopped lining up with the results. On the Artificial Analysis Intelligence Index, an independent composite of capability tests, Claude Opus 5.5 scored 58 at $20 per million output tokens, while Claude Fable 5.1 and OpenAI's GPT-6 Astra both scored 53 at $50. Claude Sonnet 5.5 scored 56 at $10 (scores compiled by AIToolsReview).
When the best model can change every few weeks, the advantage goes to whoever can switch, combine and govern models fastest. That happens in a new architectural layer: the meta-harness.
In theory, switching should be easy: change the model and carry on. In practice, the model is the smallest part of what delivery teams build. Most of the work lives in the layer around it, called the harness: the loop that calls the model, the tools it can run, the files it can touch, the memory it keeps between steps and the rules for when it stops. Claude Code, Codex and Cursor are all harnesses.
A meta-harness is a layer above your harnesses that wraps each one in a common interface, so agents, policies and tools stay portable across them.
Databricks' answer is Omnigent, a meta-harness that coordinates agents built with Claude Code, OpenAI's Codex and custom frameworks. Rather than commit to one vendor's ecosystem, teams can pick the best agent for each job and keep their context, tools and governance shared. Because the code is open source, teams can inspect it, run it on their own infrastructure and keep using it wherever Databricks takes the product.
The agent definition: one file that describes the agent, its tools and its instructions, written once and portable across harnesses.
Policies: rules for what the agent may do, enforced outside the model so no prompt can rewrite them.
Cost controls: spending tracked per session, with limits that pause the agent for approval.
Sessions: live work that teammates can watch, review and steer together.
A meta-harness turns slow, expensive decisions into fast, cheap ones:
An organization that adopts the best model in days, under controls it already trusts, will outrun one that needs a quarter to rebuild.
The cost savings can also be significant. In UC Berkeley's RouteLLM research, intelligent model routing maintained 95% of GPT-4's quality on one benchmark while reducing costs by more than 85%. Not every task needs the most expensive model, and choosing the right one can make a meaningful difference at enterprise scale.
$511,000
saved a year by routing
*A heavy-adopter scenario. The median large non-tech company uses 50 million to 500 million tokens a month, while a quarter of tech companies top 50 billion, per a Bloomberg Intelligence survey.
Putting every agent behind one layer simplifies control, but it also concentrates risk. Weigh these before adopting one:
The category is also young: Omnigent is in alpha as open source and in beta as a managed Databricks service. None of this is a reason to wait. It is a reason to treat the meta-harness like any other critical system, with redundancy, review for every policy change and only the credentials it needs.
"A meta-harness replaces our agent frameworks." It does not. LangGraph, Claude Code and your SDK agents keep doing the work; the meta-harness coordinates them from above. If someone pitches it as a rip-and-replace, they are selling something else.
"One control layer means we're governed." Only for the agents that run through it, and only for the rules it can see. The meta-harness can cap spending and gate actions, but deciding who may read which record still belongs at the data layer, in catalog permissions and MCP servers. You need both.
Close enough still counts
Four questions from real polls. Guess what share of people said yes. You score for how close you land.
What percent of consumers would trust a personal AI agent more than their best friend to make a purchase for them?
What percent of employees admit they hide how they use AI at work?
What percent of consumers want an AI agent that plans a whole trip for them, from airlines to hotels to activities?
What percent of executives say AI is tearing their company apart?
Sixteen terms, four groups of four
Find the four groups of 4. Four mistakes allowed. Run out and the board reveals itself!