Most firms think they can govern AI agents. Most can't resolve conflicts between them.

The gap nobody talks about

You've got three agents: one screening loan applications, one checking compliance and one approving payments. They all run on different models, with different update schedules. One day the screening agent says yes, the compliance agent says maybe and the payments agent says no. Who decides, and how do you prove what each of them said?

That's the question 858 executives were asked in a survey by OutSystems, and the answers reveal a problem that most firms haven't solved yet. Whilst 82% believe they can govern AI agents, only 26% can resolve conflicts between them ^[1]. The difference matters, because agents that can't agree become agents that can't be trusted.

It's like having three people on a call, each giving different instructions, and no recording of who said what. Weeks later, when someone asks why the payment went through, you're left guessing.

What agents need that dashboards don't give them

Around 71% of respondents expect applications to coordinate work across systems and teams within five years ^[1]. That means agents talking to agents, often without a human watching. When they disagree, the organisation needs more than a dashboard showing that something went wrong. It needs a record of what each agent was told to do, what it did and what it returned.

Most governance tools today track models: which version is running, how accurate it is and whether it drifts. That works for a model that gives you predictions. It doesn't work for an agent that takes actions based on instructions it pieces together from several sources, because the question shifts from "is the model working?" to "what did the agent actually do, and on whose instruction?".

The conflict is the interesting part

Conflicts between agents aren't bugs. They're features, or at least they should be. When a compliance agent disagrees with a screening agent, that disagreement is often the whole point: it surfaces a risk that one model alone would have missed. The problem is what happens next.

If you can't show which agent said what, the conflict becomes noise instead of information. The screening agent's yes and the compliance agent's maybe both disappear into a log that somebody with administrator rights could quietly edit, and six months later nobody can agree on what the agents were arguing about.

That's where the 26% figure bites. Resolving a conflict means deciding who was right, and deciding who was right means knowing what each side actually said. Without a record you can trust, the governance layer is just theatre.

What this means for you

If you're running AI agents, or planning to, here are three questions worth asking now:

1. When two agents give contradictory outputs, who decides, and how is that decision recorded?
2. Can you show, six months later, what each agent was instructed to do and what it returned?
3. If an agent acts on bad data or a misunderstood instruction, can you prove that's what happened, or will you be guessing?

The answers don't require new models. They require a process for capturing what agents do as they do it, in a way that holds up when someone questions it later.

How a digital witness handles agent conflicts

Say you're running three agents and they disagree. Here's what gets sealed: each agent's instruction (what it was asked to do), each tool call it made and each result it returned. The fingerprint is taken at each step, so if an agent later says "I never approved that payment" or "the compliance check passed", you've got the record it can't rewrite.

Anyone you share the records with—auditors, supervisors, the person affected—can check that the fingerprints match, which means the records haven't changed since they were sealed. It doesn't tell you which agent was right, or authorise the action. It shows you what happened, and that nobody has fiddled with the story since.

That's the part most dashboards miss. They tell you agents are running. A digital witness tells you what they did.

How it would work

1. Each agent sends a fingerprint of its instruction, tool call and result to WitniumChain through the MCP server as it acts.
2. The fingerprint is sealed on a chain where any later change would show, even one made by us.
3. When there's a conflict, you calculate the fingerprints again and compare them with the sealed ones. If they match, the records are unchanged.

The content stays with you. The first 2,000 witnesses a month are free. Read the full technical documentation here.

What to watch

The survey doesn't say when the 71% expect their agents to be coordinating work, but five years puts it at 2030 ^[1]. By then, the AI Act's logging requirements will have applied to high-risk systems in Annex III for nearly three years, from 2 December 2027 ^[2]. Article 12 asks for automatic logs of events over the system's lifetime ^[2], and Article 26 asks users to monitor the system's operation ^[3]. Neither says how to protect the logs, but if an agent's logs can be quietly edited, they won't hold up the day someone questions a decision.

The governance gap the survey found isn't a technical problem. It's a trust problem, and it shows up the moment two agents disagree and nobody can prove what either of them said.