Multi-agent systems

Several agents on one problem, and the coordination that costs more than it looks.

48documents on this topic
27organizations represented
2issues named
5sourced citations
0sourced statistics

The state of it

One of 6 topics within Agent orchestration.

48 documents from 27 organizations address running several agents on one problem. The engineering material and the strategy material describe two different things under the same name, and the gap is worth holding onto.

The engineering account is specific about why multi-agent works and what it costs. Anthropic's research system uses subagents that operate in parallel with their own context windows, compressing findings back to a lead agent - and reports plainly that the architecture pays mainly by getting more tokens onto the problem. That is a real mechanism and a real bill. The same account notes that multiple agents introduce new problems in coordination, evaluation and reliability.

The strategy material tends to present multi-agent as a maturity stage to progress through. Those are different claims: one is an architecture chosen for a class of problem, the other is a destination.

The issues, by agreement

How many independent organizations name each issue as a problem. An issue is only as real as the number of separate publishers that identify it, so the count is the ranking. Bars are organizations, not documents. Where the count reads ours, no publisher here states the issue and the analysis is our own.

Where they disagree

No contradictions recorded on this topic yet.

The issues in full

Each issue carries the organizations that name it, the numbers behind it, and the remedies proposed - with the concrete steps under each. Every citation points at a section of a named document, so any count here can be checked.

Issue 013 organizations name it2026 evidence

Multi-agent systems work largely by spending more tokens, and the budget rarely says so

The performance gain from running agents in parallel is substantially a gain from spending more compute on the problem. That is a legitimate trade, and it is not usually the trade the business case describes.

Anthropic is direct about the mechanism: the architecture earns its keep mainly by getting more tokens onto a problem, with subagents working separate angles in their own context windows and returning only the condensed result to a lead agent. Read as engineering, that is a clear statement of when the architecture pays - open-ended problems where the path cannot be predicted. Read as a budget line, it means the parallelism that produces the answer also multiplies the bill, and the funding request usually reflects the first and not the second.

How to fix it — 2 approaches, 5 steps

Cost the design, not the model

Estimate token spend from the architecture - number of subagents, expected turns, context per agent - before building, and compare it against the single-agent baseline.

Done when A written cost model compares the multi-agent design against a single-agent alternative on the same task, the delta is justified by a measured rather than expected accuracy gain, and per-run and per-day ceilings are set.

  1. Model the token cost of the multi-agent design against a single-agent alternative on the same task.0-30 daysCTO
  2. Require the delta to be justified by a measured accuracy gain, not an expected one.30-90 daysCIO
  3. Set per-run and per-day ceilings before the first production deployment.30-90 daysCFO

Reserve the architecture for genuinely open-ended work

Where the path can be predicted, parallel exploration is paying for search you did not need.

Done when Every multi-agent deployment in production has a written statement of why its path cannot be predicted, and anything with a knowable sequence runs as a workflow.

  1. Test whether the task has a knowable sequence. If it does, use a workflow.0-30 daysCTO
  2. Apply multi-agent designs to open-ended search and synthesis, where breadth is the point.30-90 daysCTO
The evidence — 4 documents
OrganizationDocumentPosition
AnthropicFrontier labBuilding multi-agent researchOur reading Attributes the gain largely to budget rather than cleverness: parallel subagents each hold their own context and hand a condensed result back to a lead, which is how enough tokens reach the problem.Benefits of a multi-agent system; performance analysisnames it
DeloitteConsultancyNavigating AI spend dynamicsOur reading Prescribes budget alerts, context window limits and API usage caps, controls that exist because parallel and long-running work consumes unpredictably.Spend visibility and guardrailsnames it
InfosysConsultancy · June 2026AI token economics for executivesOur reading Frames token consumption as the executive cost discipline that architecture decisions determine, which is precisely what a parallel multi-agent design decides.Token economics for executivesnames it
AccentureConsultancy · July 2025Accenture CAIO on why agentic AI is closeOur reading Presents multi-agent systems as the advanced stage requiring orchestration frameworks, agent-to-agent protocols, robust memory management and deeper enterprise integration.The multi-agent futureproposes a fix

Issue 021 organization name it

Multi-agent is being adopted as a maturity stage rather than chosen for a problem

Stage models present single agents, then multi-agent systems, as a progression to advance through. The engineering material treats multi-agent as an architecture suited to a specific class of problem, which is a different claim.

The difference matters because a maturity model implies you are behind if you have not adopted the later stage, while a problem-fit argument implies you should adopt it only where the problem has the right shape - open-ended search where the path cannot be predicted. The academic material is more sceptical still: the case that agents alone are sufficient has been questioned directly, and previous waves of agent enthusiasm are part of that argument.

How to fix it — 1 approach, 2 steps

Test for problem shape, not maturity stage

Ask whether the task is genuinely open-ended. If the path is predictable, a stage model is telling you to buy complexity you do not need.

Done when Each multi-agent approval records why the path cannot be predicted in advance, and any case where it could be is recorded as a workflow decision taken against the stage model.

  1. For each proposed multi-agent deployment, state why the path cannot be predicted in advance.0-30 daysCTO
  2. Where it can be, use a workflow and record the decision against the stage model.0-30 daysCTO
The evidence — 1 document
OrganizationDocumentPosition
AnthropicFrontier labBuilding multi-agent researchOur reading Frames multi-agent as fitted to open-ended problems where the required steps cannot be predicted in advance, rather than as a stage every organization should reach.Benefits of a multi-agent systemnames it

Who is represented

This dossier is drawn from 27 organizations working on the subject, 4 of which are cited directly in the issues above.

Consultancy — 10

Deloitte 4 Accenture 3 Infosys 1 Capgemini 3 McKinsey & Company 3 Boston Consulting Group 2 EY 1 KPMG 1 PwC 1 QuantumBlack, AI by McKinsey 1

Institution — 4

Cloud Security Alliance 2 Citi 1 FinOps Foundation 1 Moody’s 1

Academic — 1

arXiv (research) 3

Hyperscaler — 6

IBM 3 Microsoft 3 AWS 2 Google Cloud 2 Fujitsu 1 Workday 1

Frontier lab — 2

Anthropic 3 OpenAI 2

Enterprise — 1

BMW 1

Vendor — 2

Cursor 1 Palo Alto Networks 1

Other — 1

CISA 1