Agentic AI Engineering for Energy Trading: the founding cohort

Trading firms are hiring AI engineers. You already have most of what they ask for.

A London investment-bank contract advertised in November 2026 asks for a "proven track record of delivering Agentic AI solutions in production environments, or demonstratable experience building AI tools inside or outside of work setting".

The second half of that sentence is what this cohort gives you.

Agentic AI is software in which a language model decides its own next step. Each such program is called an agent.

Join the priority list

Try this first

An AI assistant sends a trade to market for a trader. The trade is executed. The reply is lost on the way back. The assistant tries again.

What happens?

Screenshot of the working system with one safeguard taken out. The provider did not reply, so the software sent the order again. The provider executed twice and the book holds one trade, so the book is wrong by 20 MW. A note explains what went wrong: asking about the order already sent is what prevents this.
From the working system, with one safeguard taken out: sold twice, booked once.

It sells twice. The firm's own book shows one trade. Nothing crashed, and every message along the way looked like success.

The book, or trading book, is the firm's record of its trades and positions.

A model that reads a trader's sentence correctly is the easy part to demonstrate. Knowing why that happened, and how to build software where it cannot, is the part firms are hiring for. Their own words in briefs from November 2026: "prototype to production", "secure", "governed", "human oversight", "audit trails".

What you build

You are given a small working system: a trader's typed request is read by a model, checked by code against a trading book's limits, approved by a second person, sent to market, and recorded. Then you change it, break it and extend it.

  • Nine short exercises, each with a test that fails until your change is right. Add a second trading rule. Let an approver refuse. Handle a reply that never comes.
  • Each safeguard removed, by you, to see what it was preventing: a trade sold twice, a trade nothing knows about, a limit broken on a stale approval.
  • One change of your own, assessed. A new approval rule, a new failure to survive, or a second provider that behaves differently. It cannot be done by following a guide.
  • A mock technical interview on your own work, including a failure you have not seen.

Two things to be plain about. The market connection in the exercises is simulated; no real order leaves your laptop. The trading book is real software, the open-source Glasshouse.

Screenshot of the working system. A trader, Dan, types: sell 20 MW from the battery at 18:00 today, limit 92. The model fills in a form with the details of the trade. Code then checks the request against the book: Dan may trade this book, the hour stays inside its limit of 100, the battery can deliver, and a trade of 25 MW or less needs no approval.
From the working system: a request read by the model and checked by code.

What it covers

Seven milestones on one small trading desk, each ending in something you show:

  1. A model turns a trader's message into a form, and you catch it making a mistake.
  2. It looks things up for itself, inside limits you set.
  3. It answers from the desk's contracts and shows the clause it used.
  4. It reads the trading book as a named person, and your tests can stop a bad release.
  5. The controlled trade: from the trader's words to a recorded trade, with approval.
  6. Several agents working together, and two firms' agents settling something neither may decide alone.
  7. All of it running in the cloud, and your mock interview.

The tools the adverts name are in it, each used on a real problem: Python, the OpenAI and Anthropic APIs, retrieval (RAG and GraphRAG), LangChain and LangGraph, MCP, A2A, evaluation, tracing, CI/CD and Azure.

  • RAG: the model searches documents before it answers. GraphRAG also follows the links between them.
  • LangChain and LangGraph: Python libraries for building agents and workflows.
  • MCP: a standard way to give a model tools and data. A2A: a standard for agents of different firms to talk.
  • Evaluation: measuring whether the answers are right. Tracing: a record of every step an agent took.
  • CI/CD: a pipeline that tests each change and then releases it. Azure: Microsoft's cloud.
Screenshot of the working system. A sale of 40 MW is above 25 MW, so a second person must approve it. Before Maya approves, another trader sells 15 MW in the same hour, and the trade would now break the limit of 100. The code checks the approval against the book as it is now, and the approval is void. Below, an exercise asks the student to write a less blunt rule and check it with a test.
From the working system: an approval goes void because the book moved.

What the market looks like

UK contract adverts that mention AI agents, counted the same way over the same six months of each year: 3 in 2024, 159 in 2025, 982 in 2026. Across those 982 the middle day rate is £560 and the top tenth pay £838 or more. (IT Jobs Watch, read 10 November 2026.)

Permanent adverts, counted the same way: 2 in 2024, 135 in 2025, 1,137 in 2026. Across those 1,137 the middle salary is £95,000 and the top tenth pay £142,500 or more. (IT Jobs Watch, read 11 November 2026.)

Those figures cover every sector. We do not have a figure for energy trading alone.

Who it is for

Developers and technically able analysts who already work in or near energy trading. You bring the years. This adds the part the new adverts ask for, in a form you can show.

It is not for people new to programming, or for anyone wanting a general introduction to AI.

What it takes

One live session a week, recorded, at a time agreed with the group. The rest is done in your own time, at your own pace, with help between sessions.

You need to be comfortable reading and changing Python, and with ordinary arithmetic about power, energy and money. You do not need machine-learning mathematics, and you never train a model.

The founding cohort

Twelve places. Included:

  • seven live sessions with Jordan Dimov
  • one personal review of your code and your design
  • one mock technical interview
  • feedback in the group throughout

It is the first run. The written guides are produced as the cohort goes, and founding members shape them.

The start date is set with the first members, and people on the priority list hear it first. The price is sent in reply to everyone who joins the list.

Lifetime members of the community have priority for a place, and what they paid for lifetime membership is taken off the price.

Who runs it

Jordan Dimov built position, profit-and-loss and regulatory reporting systems at Shell, and has put large language model (LLM) extraction and agent systems into production. He wrote the system the cohort is built around.

What we will not tell you

  • We cannot promise you a seat or a rate. You will have built, tested and defended one controlled workflow. What a firm does with that is the firm's decision.
  • This cohort does not replace years of engineering. It adds a new specialism to the experience you have.

Join the priority list for the founding cohort

Tell us where you are and where you want to be. Jordan reads every answer. If this will not get you there, he will say so.

Only your name and email are required. We use your answers to plan this cohort and to write to you about it, and for nothing else. Privacy policy.

If your employer might pay, say so on the form and we will send you a page to give your manager.