In agentic post-trade, explainability remains one of the hardest problems to solve. Agents can already read systems, draft exceptions, propose amendments and assemble workflows. What they have struggled to do – in a way that satisfies a Head of Operations facing a regulator – is leave behind an audit trail that survives scrutiny.
Duco’s launch this week of what it describes as an agentic Operations platform is, at its core, a specific architectural answer to that problem. The reconciliation specialist has unbundled its platform into close to 200 discrete capabilities, exposed through Model Context Protocol (MCP) as a “tool surface” that agents call rather than replace. Matching, rules and audit logging remain deterministic. The agent layer sits on top, orchestrating those tools and, in doing so, generating an audit point every time it makes a call.That distinction is where the launch separates itself from a wider field of post-trade vendors that have made agentic announcements over the past year.
Why monolithic agents do not work in regulated workflows
The naïve approach to agentic post-trade is to hand the model everything: data, instructions, exception logic, decisioning. Duco’s argument is that this fails on two counts simultaneously – economic and regulatory.
“You could try to do everything through the model. You could say: here’s some data, go reconcile it, figure out the exceptions, risk-assess them,” Christian Nentwich, CEO and co-founder of Duco, tells TradingTech Insight. “The problem is, if you do all that through the model, it’s incredibly expensive – you’re going to burn billions of tokens – but the other thing is, you then have no audit trail of how any of the decisions were made. By getting the agents to talk to a tool, and to another tool, and then make a decision, every time it calls a tool there’s an audit point, and something is logged.”
There is also a practical engineering constraint that pushes in the same direction. Duco stores more than a trillion records and ingests upwards of a billion records a day. “You can’t just jam that into the context window of the model,” Nentwich says. “It doesn’t work – it starts to hallucinate. Even if you just try to fit 160 tools in there, it starts hallucinating because it’s overloaded with context.”
The implication runs beyond Duco. For any post-trade vendor exposing capabilities to agents, scale alone forces a tool-surface architecture rather than a monolithic one. The interesting question becomes how that surface is designed: which capabilities are exposed, how they are described to the model, and where the deterministic boundary sits between what the agent decides and what the platform executes.The architectural argument as regulatory answer
Nentwich frames the tool-call-as-audit-point construction explicitly as the response to the explainability question.
“We need to make sure they pass audits at the end of the day,” he says. “We’re learning where people’s risk appetite is for sign-off, but we’re trying to get people away from this purist idea that the model is just going to do everything itself. We don’t think that works in financial services.”
Although several competitors have shipped MCP-accessible tools into post-trade workflows, where Duco’s launch goes further is in arguing that exposing tools is not the point – the point is that those tools take deterministic action and emit auditable evidence. On that reading, MCP is connective tissue. The architecture underneath is what matters.
The Pacesetters cohort: deliberately constrained
Duco has gone live this week with 10 clients in what it is calling the Pacesetters cohort. The composition, by Nentwich’s account, is three European global banks, a set of asset managers, several private credit funds and one broker. The firm has declined to name participants at this stage.
The cohort is small by design. “We had more demand than this – we were deliberately constraining it,” Nentwich says. “We want to learn more about how our clients want to adapt their operating model and what return on investment they’re getting. You don’t want to run the same operations processes with agentic capability like this – you really need to figure out how to reorganise yourselves.”
Early indications come from a single use case that Duco has highlighted: a client whose 15-year-old Excel macro for reconciliation configuration was rebuilt on the Duco platform in four hours, with roughly 20 minutes of agent runtime.
While that time saving is impressive in itself, a more telling detail sits beneath it: in the course of the rebuild, the agent surfaced several errors in the macro that the client had not previously known about. That is a more useful illustration of agentic capability than configuration speed alone, because it speaks to capability the human operator did not have.
The interface question
The boldest claim Nentwich makes is one that goes beyond the launch itself.
“We think ultimately people won’t log into platforms like Duco anymore,” he says. “We’re in a transition period, but there will be a world where there is no user interface – just agents running things and using bits and capabilities of platforms like Duco to get work done.”
Duco has been explicit that its standard SaaS model remains available alongside the agentic workspace. But the direction of travel Nentwich describes – platforms as substrate rather than destination, accessed by agents rather than humans – is one that has implications across the post-trade vendor landscape, not just for Duco. If correct, the competitive question for vendors shifts from interface design to tool surface design.
What is not yet known
Nentwich is candid about the limits of what the launch can demonstrate at this stage.
“There are a lot of good questions about what the actual impact of this is on daily work, and the truth is we don’t know yet,” he says. “We’re about to find out.”
This framing points to the more interesting story still to come. The architectural argument for agentic post-trade is easy to make. Whether the operating model holds up under production conditions, across a heterogeneous client base, and against the explainability standards that regulators will apply, is a question that will be answered by what the Pacesetters report when they look at their reconciliation desks over the coming months.
Subscribe to our newsletter
