Skip to main content
Viithiisys
When two agents disagree

Several agents, one outcome, and a clear answer to what happens when two of them disagree.

Agentic systems are worth building when work genuinely needs more than one specialised agent coordinating. That coordination, and its failure modes, is the whole engineering problem.

Building and running software since 2007One agent first, wherever one agent will do

One taskTwo agents
Task

The same task, given to both agents

Agent A

Answer A

Confidence: high

Agent B

Answer B

Confidence: high

Who decides?

The system picks one, usually the last or the most confident-sounding, and nothing records that a conflict occurred.

Illustrative of the disagreement case, not a client result.

Before you split it

When one agent is not enough, and when it is.

This is the wrong page for most buyers, and that is worth saying first.

  1. The work needs genuinely different specialisms

    One part reads documents, another queries a system, another drafts a response. Each needs different context and different permissions.

  2. A single agent has become an unmaintainable prompt

    It works, and nobody can change it safely, because every instruction interacts with every other one.

  3. Steps need to run in parallel and then reconcile

    Three lookups at once, then a decision from all three. Sequential handling makes the whole task as slow as the sum of its parts.

  4. Different steps need different authority

    Reading is safe, writing is not, and one agent holding both permissions is a larger risk than the work requires.

If the work is one job with one set of permissions, a single agent is simpler and easier to debug. Multi-agent architecture answers genuine specialisation, not ambition.

See AI agent development
The arrangements

Four shapes that work.

An agent
An agent
Work passing between them
Work passing between them
A checkpoint
A checkpoint
A permission boundary
A permission boundary
  1. SupervisorWorkerWorkerWorker

    A supervisor with specialised workers

    One agent decomposes the task and routes parts to agents that each do one thing well. The supervisor holds the plan; the workers hold narrow permissions.

  2. 01020304

    A pipeline with checkpoints

    Stages in sequence, each validating the previous stage's output before proceeding. Slower than parallel and far easier to debug and defend.

  3. TaskAgentAgentAgentReconcile

    Parallel retrieval with reconciliation

    Several agents gather independently and one reconciles, including the case where two return contradictory answers.

  4. AgentToolToolTool

    A single agent with tool boundaries

    Frequently the right answer

    Listed here because it is frequently the right answer. One agent, several tools, distinct permissions per tool. Less architecture, same outcome.

What only happens with several

The four failure modes specific to multi-agent systems, and what each requires.

These do not occur with a single agent, which is why multi-agent work is more than a scaled-up version of it.

  1. Disagreement

    Two agents return contradictory results.

    What it looks like in production

    The system picks one, usually the last or the most confident-sounding, and nothing records that a conflict occurred

    What it requires at design time

    An explicit reconciliation rule, plus surfacing conflicts rather than resolving them silently

  2. Loops

    Agents hand work back and forth without converging.

    Produces incidents
    What it looks like in production

    Cost and latency climb with no output. Frequently discovered on an invoice rather than in monitoring

    What it requires at design time

    Hard limits on hops and total spend per task, enforced rather than advisory

  3. Diffused accountability

    The output is wrong and no single step was.

    Produces incidents
    What it looks like in production

    Each agent behaved correctly given its input. The error is in the composition, and it is not attributable

    What it requires at design time

    Trace every step's input and output, so a wrong result can be walked backwards

  4. Permission creep

    A worker agent needs one more capability, repeatedly.

    What it looks like in production

    Over months, the narrow agents each accumulate broad access, and the isolation the design depended on is gone

    What it requires at design time

    Per-agent permission review, and treating a permission increase as a change requiring approval

The second and third produce incidents. A loop is a cost event; diffused accountability is what makes an incident impossible to explain afterwards.

How the work runs

How the work runs, step by step

This is the architecture; the agents themselves are comparatively straightforward.

  1. Try the single-agent version first

    On paper, and often in practice. If one agent with well-scoped tools can do the job, that is the answer and the engagement is smaller than expected.

  2. Decompose only where specialisms genuinely differ

    Different context, different permissions, or different failure consequences. Splitting on anything else adds coordination cost for nothing.

  3. Design the coordination contract

    What each agent receives, what it must return, and what happens when it cannot.

  4. Set the hard limits before the first run

    Maximum hops, maximum spend per task, and a stop condition. Made afterwards, these are written in response to a bill.

  5. Trace everything, from the first day

    Every step's input and output retained. Without it, a wrong answer in production cannot be explained, only apologised for.

What we can point at

Where this work has been done

Coordination and cost control are the hard parts of agentic work, and both are visible in what Viithiisys already runs in production.

Workplaces where Vizitor is in daily use
500+Workplaces where Vizitor is in daily use
Countries Vizitor runs in
15+Countries Vizitor runs in
Years of Viithiisys delivery, since 2007
19Years of Viithiisys delivery, since 2007
Countries Viithiisys has delivered across
6Countries Viithiisys has delivered across

None of these is a multi-agent engagement, and none is offered as one. They are cited as context for the two things this page says are hard, coordination and cost control.

See the case studies
What it costs to run

Cost per completed task, not cost per call, and the difference is where projects get caught.

A single agent answers once. A multi-agent system may make a dozen model calls to complete one task, and that multiplier is invisible during a prototype.

Model calls per completed task
A single agent

One model call

A multi-agent system

A dozen model calls

monthly volumecost per completed task
  1. Calls per task, not calls per agent.

    Why it is invisible early. A prototype is run by hand a few times. Nobody multiplies by monthly volume

    What to put in placeMeasure calls per completed task during the pilot, then multiply by real volume before committing

  2. Retries and re-planning.

    Why it is invisible early. A supervisor that re-plans after a worker fails can double the cost of a task that still succeeds

    What to put in placeCount retries separately. A rising retry rate is a quality signal as well as a cost one

  3. Context carried between agents.

    Why it is invisible early. Each handoff may resend accumulated context, so cost grows faster than the number of steps

    What to put in placeDecide what each agent actually needs. Passing everything is the default and the expensive one

  4. Traces and retention.

    Why it is invisible early. Storage is cheap until you keep every step of every task for a year

    What to put in placeSet a retention period deliberately, balanced against how far back an incident might need walking

  5. The loop case.

    Why it is invisible early. Rare, and disproportionately expensive when it happens

    What to put in placeHard hop and spend caps, which is the same control as the loop failure mode above

Know cost per completed task before production, not after. A system that is correct and uneconomic gets switched off, which is the same outcome as one that does not work.

Which page you want

Four pages could answer an agent question. Here is which one you want.

The titles do not make the distinction obvious, and choosing wrongly costs a first conversation. This page is the narrowest of the four.

What you keep

What you are left holding.

The architecture is the asset.

  1. The coordination contract, documented

    What each agent receives and returns, and the reconciliation rule. Portable across frameworks, which matters in a category changing this quickly.

  2. Traces you can query

    In your systems. The ability to walk a wrong answer backwards is what makes the system defensible to a customer.

  3. The limits, enforced in code

    Hop and spend caps as behaviour rather than as documentation.

  4. No framework lock-in beyond what you chose

    Orchestration frameworks in this category are young and will be replaced.

Start here

Describe the task and how many different things it needs to do.

If the answer is one thing, the single-agent page is cheaper and we will say so. If it is genuinely several, with different permissions, this is the right conversation.

FAQ

Questions we are actually asked

Pick a topic, or ask us directly. We answer every inbound within one business day.

Still have questions?

Talk to a senior engineer, not a bot.

Talk to us