How is a coherence percentage computed?

At its simplest: the count of shared facts with no open conflict divided by the total shared facts in the mission. Conflicts are detected by optimistic concurrency — a write based on a stale version of a fact is flagged rather than applied — so the numerator and denominator are both precisely known. There is no sampling or model-judged scoring.

Does high coherence mean the agents' output is correct?

No. Coherence measures agreement, not correctness — a fleet can be perfectly coherent about a shared assumption that happens to be wrong, and no conflict will ever fire. It also only counts facts in the shared store; coordination through side channels is invisible to it. Use coherence to know where to look, and the underlying trace to decide what to do.

Measuring multi-agent alignment: the coherence metric

A fleet of agents is only as useful as it is aligned. Coherence is the single number that says whether the work actually agrees with itself.

By Quordo · Published 2026-06-19 · Updated 2026-06-19 · 7 min read

Key takeaways

Coherence is the share of a mission's shared, versioned context with no open conflicts — a single mission-scoped percentage that says whether the agents' work agrees with itself.
It is computed deterministically from optimistic-concurrency conflict detection against a versioned store, with no sampling or model-judged scoring.
Coherence measures agreement, not correctness: agents can be perfectly aligned on a wrong assumption, so it is a coordination signal, never a quality guarantee.

What the coherence metric measures

Coherence is the share of a mission's shared, versioned context that has no open conflicts: the percentage of facts every agent is working from that no other agent has contradicted without resolution. A mission at 100 percent coherence means its agents are all acting on a context that currently agrees with itself; a mission at 70 percent means roughly a third of its shared facts are in dispute and someone is building on a contradiction. It is a single, mission-scoped alignment number, and it is meant to be read at a glance the way you read a build status or an uptime figure.

The reason to collapse alignment into one headline number is that a fleet of agents produces far too many individual events for a human to watch live, and you cannot manage what you cannot see at a glance. A per-mission coherence percentage gives an operator one thing to look at to answer the only question that matters before trusting the output: are these agents actually aligned, or are they confidently diverging? A low or falling coherence score is an early signal that the work is drifting apart, surfaced while it is still cheap to fix rather than after the assembled result breaks.

Crucially, coherence is a property of the shared context, not of the models. It does not ask whether any single agent is smart or whether its answer is correct in isolation. It asks whether the agents agree on the facts they are jointly responsible for. That is what makes it measurable without grading anyone's output: you are counting unresolved conflicts in a shared store, not judging quality, which is what lets the number be computed continuously and trusted as an operational signal.

Why capability benchmarks structurally cannot measure this

Every evaluation that matters today scores a single agent: how well a model plans, writes code, retrieves, or reasons in isolation. Coherence measures something those benchmarks cannot reach, because it lives in the relationships between agents rather than inside any one of them. You can assemble a fleet of individually excellent agents and still get collectively incoherent work, and no single-agent score will ever show it. The failure is not a weak model; it is two strong models acting on facts that no longer agree.

That failure mode is quiet by construction. None of the agents throws an error, none is wrong on its own terms, and each produces output that looks correct when read alone, so the divergence accumulates unseen until the pieces are supposed to fit together and do not. Because nothing fails loudly, you need a metric that watches the connective tissue directly rather than inferring health from the quality of individual answers. A capability score going up tells you the agents are smarter; it tells you nothing about whether they are pulling in the same direction.

This is why a separate alignment metric has to exist at all. Coherence is not a better capability benchmark or a quality score in disguise. It is a different axis: a coordination measurement whose entire job is to make the invisible disagreement between agents visible and countable, so that the one failure mode capability scores are blind to becomes the one number an operator can watch.

What to measure: conflicts, drift, duplication, and handoff integrity

The most direct input to coherence is the open conflict, and the cleanest way to detect one is optimistic concurrency. Each shared fact carries a version; when an agent writes back based on a version that is no longer current — because another agent changed that fact in between — the write is flagged as a conflict rather than silently clobbering the other agent's work. An unresolved conflict is a precise, machine-detectable statement that two agents disagree about a fact they both depend on. Counting open conflicts against total shared facts is the core of the coherence percentage, and it requires no model judgment to compute.

Drift is the slower cousin of conflict: agents acting on assumptions that were true earlier in the mission but have since gone stale. Where a conflict is a collision on a single fact, drift is a whole agent operating from an outdated picture, and it shows up as a cluster of writes that contradict the current context. Duplicated work is a third signal — two agents independently solving the same problem because neither knew the other had started — which both wastes spend and tends to produce two slightly different answers that then conflict. Each of these is a way the shared context stops agreeing with itself, and each can be read directly off the same versioned store rather than estimated.

Handoff integrity is the fourth dimension and the one most often lost. A handoff is the explicit transfer of work from one agent to the next, and it is precisely where intent and assumptions leak away: the receiving agent gets a task but not the full context behind it, and proceeds on a reconstruction that may not match. A coherence measure worth trusting treats handoffs as first-class events whose context is carried and verifiable, so a clean chain of handoffs counts toward alignment and a lossy one counts against it. Measure all four — conflicts, drift, duplication, and handoff integrity — and you have captured the distinct ways a multi-agent mission comes apart, each as a countable event rather than a hunch.

How a coherence percentage is computed and used

At its simplest, coherence is the count of shared facts with no open conflict divided by the total count of shared facts in the mission, expressed as a percentage. Because conflicts are detected by optimistic concurrency against a versioned store, the numerator and denominator are both things the system already knows precisely — there is no sampling, estimation, or model-judged scoring involved. That determinism is what makes the number safe to act on: a 92 percent coherence reading means a specific, countable set of facts is in dispute, and you can open the trace and see exactly which ones.

Operationally, the number is most useful as a gate and a trend rather than a vanity figure. A mission whose coherence is high and stable is one whose output a human can review at the end instead of supervising throughout; a mission whose coherence is falling is one to look at now, before more agents build on the contradiction. Teams can set a floor — refuse to ship a mission's result while coherence sits below a threshold — the same way they gate a deploy on a passing build, turning an abstract sense of alignment into an explicit, enforceable check. The unresolved conflicts behind a low score are the worklist: resolve them, the score climbs, and the resolution is recorded in the trace.

Because coherence is mission-scoped, it composes naturally with cost. The same mission that carries a coherence percentage also carries a spend figure, so an operator can see not just whether the agents agreed but what the agreement, or the failure to agree, cost to produce. A mission that fanned out to four agents, conflicted repeatedly, and burned twice the budget to reach a mediocre coherence score is a coordination pattern to redesign — and you can only see that when alignment and cost sit on the same record.

The limits of a single alignment number

Coherence measures agreement, not correctness. A fleet of agents can be perfectly coherent about a shared fact that happens to be wrong: if every agent reads and accepts the same flawed assumption, no conflict ever fires and the score reads 100 percent while the mission marches confidently off a cliff. The metric tells you the agents are aligned, which is necessary for good collective work but never sufficient. It is a coordination signal, not a quality or truth signal, and treating a high coherence score as proof of a good outcome is a category error.

The number is also only as honest as the context is complete. Coherence can only count conflicts among facts that actually live in the shared, versioned store; assumptions an agent keeps in its own head, or context passed informally outside the tracked handoff, are invisible to it. A team that routes most of its real coordination through side channels will see a flattering coherence score that reflects how little is being measured rather than how aligned the work is. The metric rewards putting shared state where it can be seen, and it is misleading in proportion to how much is kept out of view.

Finally, a single headline number compresses away detail that sometimes matters. Two missions at 85 percent can be very different — one with a handful of trivial conflicts on peripheral facts, another with a few critical conflicts at the center of the work. The percentage is the right thing to glance at, but it is a pointer into the trace, not a substitute for it; the conflicts behind the number are where the actual decisions live. Use coherence to know where to look, and use the underlying record to decide what to do.

Where Quordo fits

Coherence is the headline alignment metric on Quordo's Coordinate surface, and it is computed exactly as described: the share of a mission's shared, versioned context with no open conflicts, where conflicts are detected by optimistic concurrency rather than guessed at. Every mission gives its agents a shared, versioned context to read and write, makes handoffs between agents explicit, and records every mutation in an append-only trace spanning fifteen event types — runs, context writes, handoffs, actions, and conflicts among them. That structure is what makes the metric trustworthy: the conflicts behind a coherence score are real, ordered events you can open and inspect, not an after-the-fact reconstruction.

Because Coordinate shares one plane with Quordo's Spend surface, a mission's coherence and its dollar-cost live in the same cockpit: per-run agent cost reconciles into per-mission spend across the providers Quordo connects, so you can see whether agents agreed and what the agreement cost in one place. Coordinate speaks MCP today, with agent-to-agent (A2A) coordination on the roadmap, and it sits above whatever frameworks and vendors you already run rather than replacing them. The result is a multi-agent alignment number you can gate on, trend, and trust — backed by an auditable trace and reported alongside the cost of the work it describes.

Frequently asked questions

What is agent coherence?: Coherence is the share of a mission's shared, versioned context that has no open conflicts: the percentage of shared facts no agent has contradicted without resolution. A mission at 100% coherence means its agents are acting on a context that currently agrees with itself; at 70%, roughly a third of shared facts are in dispute.
How is a coherence percentage computed?: At its simplest: the count of shared facts with no open conflict divided by the total shared facts in the mission. Conflicts are detected by optimistic concurrency — a write based on a stale version of a fact is flagged rather than applied — so the numerator and denominator are both precisely known. There is no sampling or model-judged scoring.
Does high coherence mean the agents' output is correct?: No. Coherence measures agreement, not correctness — a fleet can be perfectly coherent about a shared assumption that happens to be wrong, and no conflict will ever fire. It also only counts facts in the shared store; coordination through side channels is invisible to it. Use coherence to know where to look, and the underlying trace to decide what to do.