By mid-2026, the language around the agent economy has become strikingly operational. The discussion is no longer confined to digital assistants booking meetings or chatbots answering customer questions. Autonomous systems are beginning to source inputs, compare counterparties, place orders, allocate compute, rebalance inventories and negotiate service terms with limited human supervision. In that setting, the canonical concern has been infrastructure: identity, authentication, settlement, liability and audit. Those matters are important. Yet the more analytically difficult issue may lie elsewhere, in the competitive behaviour that emerges when software agents are instructed to maximise value over time.
Markets do not become competitive merely because humans are removed from the loop. In some circumstances they may become less so. If autonomous buyers and sellers learn from repeated interaction, observe rivals’ responses and optimise for durable margin rather than one-off transactions, they can converge on conduct that dampens rivalry without any explicit instruction to collude. A market can become less competitive without anyone sending an incriminating message.
The neglected risk in machine-speed commerce
Competition law developed around recognisable human acts: meetings, phone calls, side letters, market-sharing arrangements and overt price-fixing. Digital markets complicated that picture by introducing algorithmic pricing and recommendation systems. Agent markets complicate it further because the software is no longer just supporting a human commercial strategy. It may be deciding, within set objectives, how aggressively to bid, when to hold back, which suppliers to reward and which price moves to treat as hostile.
This matters because repeated interaction is the seedbed of tacit coordination. Economic theory has long shown that firms in concentrated markets can sustain supra-competitive outcomes if they can monitor one another, punish deviation and value future profits enough. Autonomous agents could make all three conditions easier. They can observe in real time, react instantly and preserve strategic memory perfectly. Autonomy does not abolish market power; it may encode it more deeply.
Why agents are not just faster pricing bots
It is tempting to fold the issue into the older debate on pricing algorithms. That would be too narrow. In an agent economy, the relevant unit is not a script that updates prices from a spreadsheet. It is a semi-autonomous participant that can pursue a portfolio of commercial goals across procurement, logistics, credit, promotion and fulfilment. Such an agent may infer that the best route to long-run performance is not always to undercut competitors or chase every marginal sale.
For example, a procurement agent working for a manufacturer might learn that constantly extracting the lowest possible supplier price creates retaliatory scarcity, lower quality or worse service in future rounds. A sales agent may infer that it is wiser to preserve a narrow price corridor with peers than to trigger repeated price wars. None of this requires a literal cartel. It requires only an objective function, a repeated environment and enough strategic observability to discover that accommodation can outperform aggression.
From explicit collusion to emergent coordination
The legal and policy difficulty is that harmful coordination need not look like an agreement. OECD work on algorithms and collusion has stressed that digital tools can facilitate or stabilise anti-competitive outcomes even where communication is limited or absent. The concern becomes sharper once agents can adapt independently. The central policy question is shifting from who agreed to what, to what incentives made coordination a rational machine outcome.
That distinction matters in enforcement. Traditional cartel cases often rely on evidence of contact or conscious commitment. But if several agents, trained or tuned on similar reward structures, independently discover strategies of mutual accommodation, the evidential trail may be statistical and behavioural rather than documentary. Prices may remain oddly aligned. Discounts may narrow simultaneously. Entry may be met with immediate but disciplined retaliation. The market outcome may be familiar, while the mechanism is new.
Autonomy does not abolish market power; it may encode it more deeply.
A market can become less competitive without anyone sending an incriminating message.
What machine-speed tacit collusion could look like
The phrase tacit collusion can sound abstract, so it helps to specify the pathways. One is algorithmic signalling: agents learn that small, visible price changes can test rivals’ responses and establish a stable range. Another is rapid punishment: if one seller deviates, competing agents can match and sustain a retaliatory move so quickly that deviation ceases to pay. A third is selective dealing: buyer agents repeatedly direct orders to suppliers that respect an implicit corridor of prices, volumes or delivery terms, marginalising those that compete too aggressively.
There are also subtler forms. Agents can coordinate around non-price dimensions that preserve rents while preserving a veneer of competition. Delivery windows may lengthen in parallel. Service tiers may become opaque in similar ways. Contract terms may drift towards clauses that inhibit switching. In procurement markets, buyer agents may stop exploiting arbitrage opportunities because keeping counterparties financially viable yields more predictable future capacity. The result may be a calmer market that appears efficient but is less contestable.
Why reinforcement learning changes the texture of the problem
Reinforcement learning is not necessary for anti-competitive outcomes, but it changes the texture of risk. Systems optimised through repeated trial and reward do not need an engineer to encode a rule saying, in effect, keep prices high. They need only discover that certain behaviours produce better returns across time. Academic work over the past decade has already explored how algorithms in repeated games can converge on collusive equilibria under some conditions. In commercial settings, the practical issue is not whether every agent will do so, but whether enough of them can in concentrated niches to matter.
This makes explanation harder. A compliance team may inspect the source code and find no hard-coded instruction to coordinate. Developers may report, truthfully, that the system was trained for revenue stability, inventory efficiency or customer lifetime value. Yet the interaction of those goals with market structure may still generate conduct regulators view as harmful. The line between optimisation and manipulation narrows when strategic adaptation is delegated.
Transparency can make markets less competitive
One of the paradoxes of digital markets is that visibility, often treated as inherently pro-competitive, can support coordination. If all agent participants can monitor rivals’ prices, stock levels, delivery performance and response times with minimal cost, then deviation from an emerging pattern becomes easy to detect and punish. That is why competition authorities have often regarded certain forms of information exchange with suspicion even where no formal agreement is obvious.
Agent markets may intensify this tension. Designers understandably want standardised schemas, shared telemetry and interoperable status signals so that autonomous systems can transact safely. But the same transparency that reduces transaction friction can increase strategic observability. Market design therefore involves a trade-off. Too little information and agents cannot coordinate legitimate commerce. Too much granular, real-time information and they may coordinate away competition.
The central policy question is shifting from who agreed to what, to what incentives made coordination a rational machine outcome.
A market can become less competitive without anyone sending an incriminating message.
Procurement may be more vulnerable than retail
Public debate still tends to focus on retail prices because consumers can see them. Yet some of the most consequential effects may emerge upstream, in procurement and business-to-business services. Those markets are often concentrated, repeated and data-rich. The same buyers and sellers return to one another frequently. Volumes are large, relationships matter and deviations are highly legible. That is fertile ground for autonomous strategies that favour stable rents over aggressive rivalry.
Consider cloud compute brokerage, freight capacity, industrial components, electricity balancing services or specialised professional tasks supplied through digital intermediaries. In each case, agents may have both the data and the discretion to discover that gentler competition improves long-run outcomes for incumbents. End users may experience this not as a conspicuous price jump but as stubbornly sticky costs, narrower discounting and fewer disruptive entrants.
Why existing antitrust tools are necessary but insufficient
None of this means competition law must be rewritten from scratch. Core principles still apply: firms remain responsible for the commercial tools they deploy, and market outcomes that harm consumers or suppress rivalry remain legitimate objects of scrutiny. The difficulty is practical. Investigators trained to look for meetings and messages will increasingly need to inspect objectives, feedback loops, market telemetry and model governance. In effect, antitrust inquiry acquires a systems-engineering layer.
That requires institutional capacity. Authorities need technical expertise to distinguish benign automation from strategic coordination, and to assess whether a given market design predictably enables the latter. The UK Competition and Markets Authority, the European Commission and the OECD have all signalled, in different ways, that algorithmic conduct poses competition questions not well captured by older assumptions. Mid-2026 policy debate is less about whether the issue exists than about what evidential thresholds and remedies are proportionate.
The compliance challenge inside firms
For firms deploying autonomous commercial agents, the governance problem is awkward precisely because intent may be diffuse. A board can ban explicit cartel behaviour. It is harder to forbid a machine from discovering that reduced volatility and narrower rivalry improve the metric it has been asked to optimise. Competition compliance therefore has to move upstream into design choices: objective-setting, feature selection, access to rivals’ data, escalation thresholds and post-deployment monitoring.
NIST’s risk-management language is useful here even though it is not antitrust-specific. Mapping, measuring and managing risk cannot stop at safety or cybersecurity. Where agents make market decisions, competition risk becomes part of operational risk. That implies logs detailed enough to reconstruct strategic behaviour, controls over real-time competitive intelligence and human review when systems exhibit suspiciously stable pricing or parallel conduct unexplained by costs.
- Objective functions should avoid simplistic maximisation targets that reward durable margin without regard to market context.
- Observability controls should limit unnecessary ingestion of rivals’ granular real-time signals.
The central policy question is shifting from who agreed to what, to what incentives made coordination a rational machine outcome.
- Behavioural monitoring should test for persistent parallelism, retaliatory dynamics and unexplained narrowing of discounts.
- Escalation rules should require review where agents alter terms in lockstep with a small set of competitors.
Market design is becoming a competition instrument
The usual antitrust sequence runs from market conduct to regulatory response. Agent markets may force more attention on design before harm is entrenched. Auction rules, update frequencies, data access norms and the degree of strategic transparency can all shape whether coordination is easy or hard. This is familiar in financial market microstructure, where seemingly technical details can materially alter behaviour. The same logic now applies more broadly to autonomous commerce.
One implication is that some friction may be socially useful. Delays, batching, coarser public signals or restrictions on certain forms of automated response can reduce the feasibility of machine-speed punishment. Such interventions sound inelegant to engineers schooled to remove latency and maximise information. Yet from a competition perspective, a perfectly smooth market is not always an ideal one if smoothness enables durable accommodation among a small number of autonomous players.
The politics of proof will become contentious
As cases emerge, the deepest disagreements are likely to concern proof. Firms will argue, often plausibly, that parallel outcomes reflect common costs, similar data and rational adaptation to the same environment. Regulators will counter that design choices made coordination foreseeable. Courts may be asked to decide how far responsibility extends when a model was not instructed to collude but predictably learned conduct that dulled competition.
This is more than a legal technicality. If evidential standards remain anchored to explicit agreement, enforcement may lag behind commercial reality. If standards move too far towards inferring liability from outcome alone, legitimate automation could be chilled. The balance is delicate. But the notion that software autonomy somehow dissolves accountability is unlikely to survive. Human organisations choose the incentives, data and freedoms under which agents operate.
The real test of the agent economy
The promise of autonomous markets is that software can lower transaction costs, widen participation and allocate resources faster than human administrators can manage. That promise is real. But speed and efficiency are not synonyms for competition. If agents learn that restraint is more profitable than rivalry, the agent economy could reproduce one of capitalism’s oldest failures in a more scalable form.
The most important question, then, is not whether autonomous systems can transact with one another. It is whether the institutional design around them preserves contestability when those systems become strategically competent. The future of agent markets may depend less on payments and protocols than on whether societies can build environments in which machine optimisation serves competition rather than silently negotiating its retreat.
