In our last article, we explored how AI agents could transform commodity trading, increasing trading volumes, expanding market participation and reorganizing trading desks around portfolios of risks rather than individual commodities.

That raises a practical question.

Where will these agents actually trade?

The instinctive answer might be the electronic order book. Give an AI agent access to a trading API, connect it to an exchange, and let it execute transactions.

For highly liquid, standardized financial markets, that may be largely sufficient. Trading APIs, automated execution, matching engines and credit controls are already well established.

But much of the commodity trading world operates very differently.

Trading happens through conversations with brokers and counterparties, instant messaging, requests for quotes (RFQs), negotiated bilateral transactions, tenders for physical cargoes, and auctions that range from a single seller to complex market-clearing mechanisms.

These formats exist for good commercial reasons. They are particularly important in less liquid, fragmented markets where finding a suitable counterparty can be harder than determining an acceptable price.

If AI agents are going to transform commodity trading, they will need to operate across this entire ecosystem, not simply connect to an order book.

How commodities actually trade

Continuous central limit order books work extremely well when products are standardized, participation is broad, and there is enough liquidity to maintain executable bids and offers.

Many, and arguably most, commodity markets don't have those characteristics. Beyond the order book, four trading formats are particularly important.

Instant messaging and bilateral negotiation. Much OTC trading still occurs through conversations with brokers and counterparties, often on specialized messaging platforms. These let traders communicate interest discreetly, explore available liquidity and negotiate terms that don't fit neatly into an order book.

Request for quote (RFQ) platforms. RFQ systems let a participant distribute a trading interest to a selected group of counterparties, who decide whether to respond and at what price. Some support multiple rounds of negotiation and counteroffers.

One-to-many tenders. A producer or consumer offers or seeks a defined quantity and invites competing bids or offers. Tenders are common for LNG cargoes, crude oil parcels, natural gas supply arrangements and blocks of environmental credits.

Market-clearing auctions. More sophisticated auctions evaluate bids and offers from many participants at once, sometimes across multiple products, locations and delivery periods. Electricity markets and government emissions allowance auctions are familiar examples.

Each format addresses needs a conventional order book may not: thin liquidity, controlled disclosure, non-standardized products and the coordinated allocation of large quantities.

In many physical commodity markets, a screen displays only a small portion of the available trading interest. The rest exists in relationships, conversations, private inquiries and scheduled market events.

Automation hasn't reached every market equally

As we noted in our prior article, algorithmic trading has had its greatest success in liquid, standardized financial markets, with far less penetration in physical and OTC commodity markets. That's less about reluctance to adopt technology than about trading processes that are fundamentally different.

Technology has improved parts of those processes. Messaging systems capture conversations and help generate trade records. RFQ platforms distribute inquiries and collect responses electronically. Auctions use sophisticated algorithms to determine clearing prices and allocations.

But in most cases, humans still initiate transactions, identify counterparties, interpret commercial information, negotiate and make the final decisions.

We've automated many of the mechanisms used to trade commodities. AI agents could begin automating the commercial activity that takes place within and between them.

Could agents operate across the entire trading ecosystem?

We think they could, although different mechanisms will require different levels of capability.

Consider an industrial company that needs to buy a total of 200,000 GJ of natural gas for delivery next month.

Today, a human trader might check displayed bids and offers, call a broker, send an RFQ to several approved counterparties, or wait for a scheduled supply auction.

An AI agent could evaluate those same alternatives.

If there is enough liquidity on an order book, it could execute immediately. If the quantity is too large, it might send an RFQ to selected counterparties to reduce market impact. If the product involves non-standard delivery conditions, it could start a bilateral negotiation. If a relevant auction is scheduled, it might decide that waiting offers the best outcome.

It could also divide the volume across several execution methods, managing the total position as transactions are completed.

The agent isn't attached to a particular venue or execution mechanism. It's attached to the company's commercial objective and chooses how best to achieve it.

That's a significant change. A traditional trading algorithm is usually designed around a particular market, strategy or execution mechanism. An agent could coordinate several trading systems under one commercial strategy and risk framework.

Messaging may be the most natural bridge

Much physical commodity trading still happens through messaging between known counterparties, which makes it a natural entry point for agents.

Imagine a broker sending an inquiry about available supply for next month at AECO. An authorized agent could interpret the message, identify the product and terms, assess the price against market conditions, check credit and portfolio exposure, and prepare a response. With sufficient authority, it might continue negotiating without human intervention.

The challenge is that trading conversations rely on shorthand, context and assumptions that experienced participants understand but rarely state. An agent could misread whether an indication is firm, whether a quantity is daily or total, or which delivery and credit terms apply.

There is also a security issue. An agent reading messages from external parties must never treat instructions embedded in those messages as authority to change its own trading behaviour or bypass controls.

For these reasons, we expect messaging to evolve toward a hybrid model. Agents may interpret natural-language inquiries and communicate with human traders, but executable terms should ultimately be captured in structured, machine-readable form.

Humans and agents can keep interacting naturally, while binding commitments become unambiguous.

RFQ platforms could lead the transition

Of the existing formats, RFQ systems may be the best positioned for early agent-to-agent trading. They already provide much of the required structure: a defined product, quantity, delivery period, selected counterparties, response window and execution procedure.

Imagine a buyer's agent sending an RFQ for next month's natural gas to five approved counterparties. Some recipients are human traders. Others are agents representing producers or merchants.

The responding agents evaluate the inquiry against their own supply positions, pricing models and commercial limits, then submit firm offers, possibly followed by counteroffers. The buyer's agent selects the most attractive eligible transaction.

The platform enforces the RFQ rules, validates each participant's authority and credit, and records the trade.

Nothing about that process requires the marketplace itself to become intelligent. The agents make the commercial decisions. The platform provides a trusted environment where those decisions become transactions.

And because RFQ systems already control who receives an inquiry, they naturally support the selective disclosure that remains important in physical commodity trading.

Auctions are another important opportunity

Auctions could also suit autonomous participants well.

In a one-to-many tender, an agent might evaluate an offered LNG cargo, determine a bid based on shipping and destination economics, and submit it within the auction rules. In a multi-round auction, it could adjust its strategy as prices and quantities change. In complex market-clearing auctions, optimization models could construct bids spanning multiple products, locations, delivery periods and operational constraints.

The distinction between the participant's agent and the auction operator matters here. The clearing algorithm determines the market outcome according to established rules. The agent decides how its participant should engage with the auction.

That could make auctions more accessible to smaller firms that lack the resources to analyze complex bidding opportunities continuously. It would also let participants connect auction decisions to the rest of their commodity portfolios, rather than bidding in isolation.

Why an API alone won't be enough

An API lets software communicate with a trading platform. Connecting an agent is only the beginning.

A marketplace that supports autonomous participants must also establish who the agent represents, what it is authorized to do and when its actions create a binding commitment. We see five requirements.

1. Agent identity and delegated authority. Every agent must be tied to an identifiable legal entity and operate within a defined mandate covering products, quantities, counterparties and activities, with permissions that can be revoked immediately.

2. Independent risk and credit controls. Agents must operate within limits they cannot modify themselves. Controls must also address multiple agents from the same firm trading against one another or collectively building excessive exposure.

3. Machine-readable products and contracts. Delivery points, contract periods, product specifications, eligibility requirements and operational obligations need to be represented in a form agents can interpret reliably.

4. Auditability and surveillance. Every message, inquiry, quote, negotiation step, approval and execution should be attributable to the participant and, where appropriate, the specific agent responsible.

5. Testing and certification. Agents should be tested against a market's rules and operational requirements before receiving live trading authority. Existing automated trading controls are a starting point, but negotiation workflows introduce new risks.

None of these capabilities is entirely unfamiliar to electronic commodity markets. The challenge is building a consistent framework that supports autonomous participation across multiple trading formats.

The missing piece: structured agent-to-agent negotiation

We think this could become one of the most important infrastructure developments of all.

An order book operates by established rules for submitting and matching orders. An informal conversation between traders has far more flexibility, and far more ambiguity. Agent-to-agent negotiation needs something in between.

Picture a marketplace where agents exchange structured commercial messages. A buyer requests a price for a defined product and quantity. A seller responds with an offer, explicitly marked as firm or indicative. The buyer counters. The seller accepts, rejects or modifies the terms.

Each message has an identifiable sender, defined commercial terms and, where applicable, an expiry time. Before any binding transaction is recorded, the marketplace verifies permissions, credit, product eligibility and execution conditions.

The agents remain free to set their commercial strategies and negotiate prices. The marketplace controls the process and ensures the resulting transaction is unambiguous.

Some of these capabilities already exist in electronic RFQ and negotiation systems, and standards such as FIX provide useful foundations. The opportunity is to extend and harmonize them for autonomous participants, including physical markets with more complex contractual requirements.

That approach is far more reliable than letting agents exchange unrestricted text and decide for themselves whether a transaction has occurred.

The agents can provide the intelligence. The marketplace should provide the rules.

What about liquidity that isn't displayed anywhere?

This may be the most commercially interesting opportunity of all.

An order book can only match the bids and offers submitted to it. In less liquid commodity markets, considerable trading interest never reaches a visible order book.

A producer might be willing to sell additional gas at the right price without advertising it. An industrial consumer might buy environmental credits opportunistically without keeping a standing bid. A company with a parcel of certified natural gas might be able to satisfy several buyers seeking particular certification standards or emissions attributes, none of whom are talking to that producer.

Today, brokers create much of their value by knowing where these interests exist and bringing counterparties together. Agents could make parts of that process continuous and economical.

Imagine a marketplace that supports private expressions of trading interest. An agent indicates that its company may sell a product, subject to price, eligibility, credit and other conditions. The indication isn't a firm order and isn't visible to the whole market. Approved counterparties or their agents can discover relevant interests and start a negotiation, which then proceeds through a structured RFQ or another execution mechanism.

That points to a valuable addition to electronic commodity marketplaces:

A permissioned liquidity-discovery service operating alongside conventional order books, RFQs and auctions.

Such a service would need carefully designed disclosure and surveillance controls. Participants must stay in control of which interests are revealed, to whom and under what conditions.

But it would address one of the most persistent problems in thin commodity markets: buyers and sellers who exist without knowing about one another.

Why less liquid markets may have the most to gain

In liquid financial futures markets, agents will enter an environment where sophisticated automation already exists. Many of the obvious opportunities have already attracted heavy technology investment.

Fragmented physical and environmental markets are different. There, transaction costs can discourage participation entirely. Finding a counterparty, evaluating terms, negotiating, confirming credit and processing a small transaction may take more human effort than the opportunity justifies.

Agents could substantially reduce those costs.

Consider Alberta's TIER compliance credit markets, regional natural gas hubs or emerging certified gas markets. Trading can be intermittent, products may have specific eligibility or delivery requirements, and participants may have little incentive to watch prices continuously.

A company with surplus compliance credits could authorize an agent to monitor potential buyers, evaluate their requirements and negotiate within approved limits. An industrial consumer could evaluate supply opportunities continuously rather than only during periodic procurement exercises.

The result could be more transactions, more participants and greater market transparency.

There is an important caution. Thin markets are also more susceptible to price disturbances, information leakage and concentrated trading behaviour. Strong surveillance and independently enforced controls would be essential.

The markets that have traditionally depended most on human conversations may be the ones that benefit most when AI agents can join those conversations.

What should commodity trading infrastructure providers be doing now?

Much of the foundation already exists. Electronic order books, RFQ platforms, broker messaging systems, auction engines, confirmation services and risk-management infrastructure have developed over decades.

The priority isn't a new exchange that replaces today's trading formats. It's a new layer that lets agents work across them: identity, authorization, structured negotiation and liquidity discovery.

We would focus on five priorities.

1. Make market information, product definitions and contract rules machine-readable. Agents need consistent, reliable information to evaluate transactions across trading mechanisms.

2. Develop identity, authorization and risk frameworks for agents. Participants must be able to delegate limited trading authority without surrendering control of their market exposures.

3. Expand structured RFQ and negotiation services. These may be the most practical bridge between today's human-directed OTC markets and autonomous agent-to-agent transactions.

4. Explore permissioned liquidity discovery. Helping agents find eligible trading interest that isn't publicly displayed could be especially valuable in fragmented physical and environmental markets.

5. Support multiple trading formats through a consistent framework. An agent should be able to trade on order books, negotiate with brokers, send RFQs and participate in auctions without separate approaches to identity, risk and transaction records.

This last point may matter most. The next generation of commodity trading technology doesn't need to replace the formats we have today. It needs to let intelligent software participate effectively across all of them.

Beyond the order book

The first wave of electronic commodity trading brought enormous efficiency to standardized markets through order books and automated matching. But much of the physical commodity world still relies on negotiated transactions, broker relationships, RFQs, tenders and auctions.

AI agents could extend automation into those markets, not by forcing everything into a continuous order book, but by working within the trading structures that already exist.

In some markets, agents will execute against displayed liquidity. In others, they will negotiate bilaterally, respond to RFQs or compete in auctions. And some of the most valuable trades they find may never have appeared on a screen at all.

The next great innovation in electronic commodity trading may not be a smarter matching engine. It may be a marketplace that lets intelligent agents discover liquidity, negotiate and transact across the full range of trading mechanisms.

The first wave of electronic trading brought the order book to commodity markets.

The next wave may bring agents to everything else.