{
  "schemaVersion": 1,
  "id": "atlas-lesson-multi-agent-coordination-cost",
  "title": "Coordination can add work",
  "category": "learning",
  "format": "Atlas lesson",
  "description": "Parallel critical path, duplicated effort, state merge, correlated error and budgets.",
  "canonicalUrl": "https://semiaifoundry.com/learn/atlas/multi-agent-coordination-cost/",
  "publishedDate": null,
  "revisionDate": "2026-10-07",
  "author": null,
  "topic": "Delegated tasks and coordination",
  "series": null,
  "license": null,
  "rightsUrl": "https://semiaifoundry.com/ai-atlas/licenses.html",
  "version": "1.3.0",
  "status": "Published educational lesson",
  "evidenceScope": "This is a worked explanation, not an executing external platform, model or robot. The quantities and fixtures are authored. Source scope is recorded in the automation learning review manifest; no source benchmark or private system behavior is claimed. The product P(A)P(B) is valid only under independence. The lesson’s probabilities and durations are synthetic; no actual agent trials are reported.",
  "contentCoverage": "published-lesson",
  "artifactUrl": "https://semiaifoundry.com/ai-atlas/lesson-data/multi-agent-coordination-cost-f1694811234a9de6.json",
  "interactiveUrl": "https://semiaifoundry.com/ai-atlas/#multi-agent-coordination-cost",
  "related": [],
  "supersedes": [],
  "supersededBy": [],
  "contentHash": "cccc7cc52e3d755cfdb4f8b6e010e5cf32770c3525a40cd04bfe8f29d22c3fee",
  "contentLength": 6339,
  "jsonUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-multi-agent-coordination-cost.json",
  "markdownUrl": "https://semiaifoundry.com/agents/publications/atlas-lesson-multi-agent-coordination-cost.md",
  "sections": [
    {
      "id": "section-1",
      "title": "Coordination can add work",
      "start": 49,
      "end": 445
    },
    {
      "id": "section-2",
      "title": "The mechanism",
      "start": 445,
      "end": 1750
    },
    {
      "id": "section-3",
      "title": "Science and assumptions",
      "start": 1750,
      "end": 1778
    },
    {
      "id": "section-4",
      "title": "Coordination can add work · explicit relationship",
      "start": 1778,
      "end": 2896
    },
    {
      "id": "section-5",
      "title": "Engineering example",
      "start": 2896,
      "end": 3468
    },
    {
      "id": "section-6",
      "title": "Distinctions and tradeoffs",
      "start": 3468,
      "end": 3816
    },
    {
      "id": "section-7",
      "title": "Specifications and interfaces",
      "start": 3816,
      "end": 4249
    },
    {
      "id": "section-8",
      "title": "Interfaces",
      "start": 4249,
      "end": 4522
    },
    {
      "id": "section-9",
      "title": "Check your understanding",
      "start": 4522,
      "end": 4551
    },
    {
      "id": "section-10",
      "title": "Two required independent subchecks each succeed with probability 0.9. What is the probability both succeed?",
      "start": 4551,
      "end": 4811
    },
    {
      "id": "section-11",
      "title": "Evidence and scope",
      "start": 4811,
      "end": 5370
    },
    {
      "id": "section-12",
      "title": "Sources",
      "start": 5370,
      "end": 5881
    },
    {
      "id": "section-13",
      "title": "Continue exploring",
      "start": 5881,
      "end": 5904
    },
    {
      "id": "section-14",
      "title": "Foundations",
      "start": 5904,
      "end": 6149
    },
    {
      "id": "section-15",
      "title": "Related lessons",
      "start": 6149,
      "end": 6339
    }
  ],
  "citations": [
    {
      "title": "Explore in Atlas",
      "url": "https://semiaifoundry.com/ai-atlas/#multi-agent-coordination-cost"
    },
    {
      "title": "LangGraph workflows and agents",
      "url": "https://docs.langchain.com/oss/python/langgraph/workflows-agents"
    },
    {
      "title": "A2A specification",
      "url": "https://a2a-protocol.org/latest/specification/"
    },
    {
      "title": "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains",
      "url": "https://arxiv.org/abs/2406.12045"
    },
    {
      "title": "Published lesson data",
      "url": "https://semiaifoundry.com/ai-atlas/lesson-data/multi-agent-coordination-cost-f1694811234a9de6.json"
    },
    {
      "title": "Atlas releases and licensing",
      "url": "https://semiaifoundry.com/ai-atlas/licenses.html"
    },
    {
      "title": "Traces connect cost, latency and results",
      "url": "https://semiaifoundry.com/learn/atlas/telemetry-traces-evaluation/"
    },
    {
      "title": "Delegated tasks, messages and artifacts",
      "url": "https://semiaifoundry.com/learn/atlas/delegated-agent-protocols/"
    },
    {
      "title": "Tools, memory & agent loop",
      "url": "https://semiaifoundry.com/learn/atlas/tool-loop/"
    },
    {
      "title": "Evaluation & the AGI frontier",
      "url": "https://semiaifoundry.com/learn/atlas/evaluation/"
    }
  ],
  "recordHash": "ee8e659c297718b053b9a4913e49ac14857adb1f2d2232e751c4e09b0ca00672",
  "contentFormat": "text/markdown",
  "content": "Delegated tasks and coordination · Atlas v1.3.0\n\n# Coordination can add work\n\nParallel critical path, duplicated effort, state merge, correlated error and budgets.\n\n[Explore in Atlas](https://semiaifoundry.com/ai-atlas/#multi-agent-coordination-cost)Logical state, task evidence and declared units; no physical assembly scale\n\nLearning objectiveSeparate parallel latency, total work and dependence assumptions in a same-task worker comparison.\n\n## The mechanism\n\nMore workers can expose parallelism, but they can also repeat work, disagree about ownership or share the same wrong premise. A useful comparison holds the task, available evidence and evaluation contract fixed. It measures elapsed dependency time separately from aggregate work and communication. A join that requires both subresults inherits both failure risks; a vote does not automatically make correlated evidence independent. The source of each result matters more than the number of participants repeating it. A fixed parallel decomposition and a dynamically delegated workflow are therefore different execution designs, not guaranteed quality levels. The incident engine currently has one worker; the examples here derive coordination consequences without claiming an executing multi-agent system or a measured model reliability improvement.\n\n1.  The shared task is decomposed into subresults with explicit ownership and a defined join condition.\n2.  Independent work can overlap, but joins, coordination and conflicts add latency or repeated effort.\n3.  Repeated-run evaluation uses the same final task oracle for single-worker and multi-worker designs.\n4.  Reliability calculations state any independence assumption; common sources or shared failure causes can invalidate it.\n\n## Science and assumptions\n\n### Coordination can add work · explicit relationship\n\n$$\nP(A\\cap B)=P(A)P(B\\mid A)\n$$\n\nThe chance that both required subchecks succeed equals the first success probability times the second conditioned on it.\n\nSymbols and units\n\nA, B\n\nSuccess events for the two required subchecks\n\nevents\n\nP(A)\n\nProbability of success for the first subcheck\n\n0–1\n\nP(B|A)\n\nConditional success probability for the second\n\n0–1\n\nTwo independent checks take 6 s and 9 s. A required 2 s merge gives ideal parallel elapsed time 11 s, while summed work is 17 s. If both checks are necessary and each succeeds with probability 0.9 independently, joint success is 0.81. If their failures are perfectly correlated, their joint success is 0.9 instead, and agreement supplies no additional independent evidence. These are hypothetical probabilities, not measured agents. Adding a worker can shorten a critical path without improving the probability that the completed task is correct.\n\nAssumptions The product P(A)P(B) is valid only under independence. The lesson’s probabilities and durations are synthetic; no actual agent trials are reported.\n\n## Engineering example\n\nTwo independent checks take 6 s and 9 s. A required 2 s merge gives ideal parallel elapsed time 11 s, while summed work is 17 s. If both checks are necessary and each succeeds with probability 0.9 independently, joint success is 0.81. If their failures are perfectly correlated, their joint success is 0.9 instead, and agreement supplies no additional independent evidence. These are hypothetical probabilities, not measured agents. Adding a worker can shorten a critical path without improving the probability that the completed task is correct.\n\n## Distinctions and tradeoffs\n\nTwo agreeing workers are not automatically two independent sources of evidence.\n\nCoordination can reduce waiting on independent work while adding merge cost, duplicate effort and shared failure modes.\n\nAgent teams are engineering systems whose benefits must be evaluated against the same task and resource boundary.\n\n## Specifications and interfaces\n\nMechanism boundary\n\nWhen does another worker help or hurt?\n\nThe product P(A)P(B) is valid only under independence. The lesson’s probabilities and durations are synthetic; no actual agent trials are reported.\n\nEvidence type\n\nAuthored worked example\n\nWorked explanation with a bounded numeric or state example. This lesson does not claim an executing implementation of an external model or platform.\n\n### Interfaces\n\nDeclared input and state Input\n\nThe shared task is decomposed into subresults with explicit ownership and a defined join condition.\n\nChecked outcome Output\n\nSeparate parallel latency, total work and dependence assumptions in a same-task worker comparison.\n\n## Check your understanding\n\n### Two required independent subchecks each succeed with probability 0.9. What is the probability both succeed?\n\n1.  0.99\n2.  0.9\n3.  0.81\n\nAnswer and reasoning\n\n0.81\n\nUnder the explicitly stated independence assumption the probabilities multiply: 0.9 × 0.9.\n\n## Evidence and scope\n\nThis is a worked explanation, not an executing external platform, model or robot. The quantities and fixtures are authored. Source scope is recorded in the automation learning review manifest; no source benchmark or private system behavior is claimed. The product P(A)P(B) is valid only under independence. The lesson’s probabilities and durations are synthetic; no actual agent trials are reported.\n\nThe worked example explains this mechanism. General numerical labs are a reference; they do not execute this lesson’s full mechanism.\n\n## Sources\n\n-   [LangGraph workflows and agents](https://docs.langchain.com/oss/python/langgraph/workflows-agents)\n-   [A2A specification](https://a2a-protocol.org/latest/specification/)\n-   [τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains](https://arxiv.org/abs/2406.12045)\n\n[Published lesson data](https://semiaifoundry.com/ai-atlas/lesson-data/multi-agent-coordination-cost-f1694811234a9de6.json) · [Atlas releases and licensing](https://semiaifoundry.com/ai-atlas/licenses.html)\n\n## Continue exploring\n\n### Foundations\n\n-   [Traces connect cost, latency and results](https://semiaifoundry.com/learn/atlas/telemetry-traces-evaluation/)\n-   [Delegated tasks, messages and artifacts](https://semiaifoundry.com/learn/atlas/delegated-agent-protocols/)\n\n### Related lessons\n\n-   [Tools, memory & agent loop](https://semiaifoundry.com/learn/atlas/tool-loop/)\n-   [Evaluation & the AGI frontier](https://semiaifoundry.com/learn/atlas/evaluation/)"
}