Chapter 04 — Agent cards, not chat windows

Colleagues —

If you only read the July summit coverage, "agentic frameworks" sounded like a bullet on a slide. It is not. The executive order asked for AI agents that explore design spaces, evaluate experimental outcomes, and automate workflows. ModCon has since published the metadata contract. Several Phase I teams named the agent pattern in the project title. This issue is the stack, with names.

December is also the Factory's cash date. Agents as catalogued services is the Genesis thesis. It is also how I will sell a VA: a card, not a chat window bolted onto MLS. First paid client by the 31st.

The contract: agent cards, skills, and a catalog

DOE's ModCon page is the primary source. An agent card is a structured description that makes an agent discoverable, reusable, and interoperable. It documents: what the agent does; who provides it; how to invoke it; skills and tools; inputs and outputs; authentication; runtime; intended uses; limitations; safety. In Genesis, cards are how other agents decide whether a peer is appropriate for a task.

Alongside the card:

Genesis Skills — a collection of Claude Code skills for scientific discovery and engineering workflows. DOE says they follow the Agent Skills open standard, so they are meant to be portable across tools, not locked to one vendor chat product.
Data cards and model cards — the same idea applied to datasets and weights. Every dataset, "regardless of size, sensitivity, or publication state," is supposed to have a card.

ModCon is explicit about the split of labor: ModCon is the scientific AI capability layer; AmSC is the secure federated platform that delivers it. Model teams onboard once on AmSC; identity, data, compute, and inference are platform services. Teams can contribute agents, models, workflows, and tools through a CI/CD and registration path.

That is a platform bet. If it works, a fusion materials agent at ORNL can call a data-broker service at Argonne without a six-month MOU rewrite. If it fails, you get 278 disconnected copilots.

Agents that hypothesise: Brookhaven's MARS

Brookhaven National Laboratory led seven Phase I awards. One of them is the cleanest statement of the scientific-agent thesis I have seen in this round:

MARS: scaling Multi-Agent Reinforcement learning for Scientific hypothesis generation. BNL's description: connect multiple AI agents, each trained on complex scientific tasks such as reading literature and analyzing data, to a network that will generate better research hypotheses than today's models can.

That is not "summarize this PDF." It is a multi-agent RL system whose output is a hypothesis, scored against other agents. The same BNL slate includes an agentic-AI-driven cavity-coupled cold-atom quantum sensing platform that tunes and stabilizes sensors in real time, and Cross-Domain Scientific Reasoning through Composable Foundation Models — connecting the new Genesis models "into one continuously improving reasoning system."

If you care about whether agents can do science rather than science-adjacent text, MARS is the project to follow. Phase I is still a demonstration round. Demand the evaluation: against what human baseline, on which corpus, with which false-positive rate.

Brookhaven Lab Awarded 7 Genesis Mission Projects

Agents that run the experiment: accelerators, neutrinos, nuclear

Fermilab's collaboration list is a field guide to agentic work already sitting on real instruments:

University of South Carolina — agentic workflows for the expedited search and discovery of charged lepton flavor violation in Mu2e.
University of Alabama — AI agents for high-energy physics simulations and analysis operations.
Duke — AI-accelerated systems to detect, identify, classify, and communicate supernova events in real time for DUNE.
Fermilab-led — AI/ML resonance control for superconducting RF cavities (PIP-II, LCLS-SC, EIC, FRIB, ATLAS). Anna Grassellino's quote is the operator's version: more efficient, reliable, and autonomous accelerators, higher scientific performance, lower operational complexity.

Prometheus, the $60 million Phase II nuclear award, is the heavy-industry version of the same pattern. DOE's challenges PDF calls for explainable AI: surrogate models, agentic workflows, autonomous labs, and digital twins, with humans in the loop for design, license, manufacture, construct, and operate. World Nuclear News, citing INL, uses the same "human-in-the-loop workflows" language. This is not an unsupervised reactor. It is an attempt to put agents on the document, design, and operations pile that currently burns calendar.

The manufacturing challenge in the same PDF is even more direct: "Recent advancements in agentic and generative artificial intelligence" to navigate multi-scale systems and drive digital twins with human-in-the-loop decision support.

Open weights as the public agent substrate

Genesis-Science-1, announced August 7 with Arcee, is DOE's attempt to put a science-specific open-weight model under this agent layer. The press release lists the downstream uses in so many words: "lab assistants, simulation surrogates, scientific copilots." First contribution windows closed in August; additional deadlines are rolling, expected every three months. December is the right month to ask whether the second window opened on that cadence, and who is fine-tuning domain adapters.

I will not pretend a general open-weight model is an autonomous scientist. Combined with ModCon cards and AmSC tool access, it is the thing a university team can actually fork.

U.S. Department of Energy Launches the Genesis Open Models Initiative – Apply Now!

A calendar item for teams already inside

Office of Science FOA DE-FOA-0003612 (The Genesis Mission: Transforming Science and Energy with AI) still shows a close date of Thursday, December 17, 2026 — specifically, the deadline for Phase II applications resulting from FY26 Phase I awards. If you are on a spring Phase I team and you intend to scale, that date is the gate. Sample Other Transaction and project agreements are posted on the same page.

Do not confuse this with the OTC/ConnectWerx small-business round. Different door, different lawyers.

The AI Automation Factory

agents as catalogued services — and the cash date

If you only read the summit coverage, "agentic" sounded like a bullet on a slide. It is not. Genesis treats agents as catalogued, invokable services with cards: what it does, who provides it, how to invoke, skills, auth, limitations, safety.

That is the Factory contract. A VA for a broker is not a chat window bolted onto MLS. It is a carded service: missed-call catcher; text-back; showing scheduler; file chaser. Inputs, outputs, human checkpoint, kill switch. If I cannot write the card, I do not ship the agent.

December 17 is DOE's Phase II-from-Phase-I gate for teams already inside. December 31 is mine: first paid client, cash in the account. Sudipa is the first workflow. The offer is $2,500 setup + $2,000/month. If you run a sales or broker desk and one loop is eating the week, that is the ask. One ask. Not a spray.

Worth clicking

ModCon: Transformational AI and Data — agent / data / model card templates. Read before you build another one-off agent.
Brookhaven Lab Awarded 7 Genesis Mission Projects — MARS and composable foundation models, in the lab's own words.
Launching the Genesis Mission — Section 3(a)(ii) is the agent mandate.
DE-FOA-0003612 — December 17 Phase II-from-Phase-I deadline.
How the Genesis Mission's American Science Cloud Advances Innovation — the orchestration layer agents will actually call.

January: the machines. Frontier is still the exascale workhorse. Lux was the 2026 AI injection. Discovery is the 2028 bet. The software story only matters if those boxes stay full.

The AI Automation Factory. Virtual assistants for sales and broker desks. First workflow: a real-estate broker/operator (Sudipa). Offer: $2,500 setup + $2,000/month. First paid client / cash by 31 Dec 2026. 2027: this newsletter + one weekly ask. No spray.

Reply if one loop on your desk is eating the week. deep-heap.com

Deep Datta
Deep Heap

Intelligence has a substrate.

I write this for people who already know the labs, the machines, and the constraints — and for operators who need a virtual assistant that actually runs a desk. If I cannot verify a source, I leave it out.