{"course":{"slug":"agents-and-automation","title":"Agents and Automation","url":"https://addaly.com/learn/agents-and-automation","level":"intermediate","track":"build"},"lesson":{"slug":"multi-agent-or-not","title":"When a second agent helps, and when it just costs more","url":"https://addaly.com/learn/agents-and-automation/multi-agent-or-not","position":20,"of":72,"minutes":9,"answer":"A subagent buys you a separate context window and nothing else; pay for it only when isolation or genuine parallelism is the problem you have.","keyIdea":"A subagent buys you a separate context window and nothing else; pay for it only when isolation or genuine parallelism is the problem you have."},"body":{"format":"markdown","text":"## What a subagent actually is\n\nStrip the diagrams away. A subagent is a fresh message list, with its own tools, its own budget and its own loop, whose final message comes back to the parent as a tool result.\n\nThat is the entire mechanism. There is no new capability, no emergent teamwork, no division of expertise beyond the two prompts you wrote. A \"team of five specialist agents\" is five prompts and five while loops.\n\nSaying it plainly matters, because the diagrams imply an organisation and the code is a function call.\n\n## The one thing it buys\n\n**Context isolation.** A parent needs five suppliers researched. Done in one context, that is five sets of search results (perhaps 150,000 tokens) dragged through every remaining step. Done as five subagents, each burns its 30,000 tokens in its own window and returns 400 words. The parent's transcript grows by 2,000 tokens instead of 150,000.\n\nThat is real, it is large, and it is the main honest reason to split. Genuine parallelism is the second: five independent lookups can run at once, and wall-clock time drops accordingly.\n\n## What it costs\n\n**Tokens multiply.** Each subagent re-sends its own growing transcript, and the parent still pays for orchestration. One published account of a multi-agent research system reported it using on the order of fifteen times the tokens of an ordinary chat interaction. Budget for a multiple, not a margin.\n\n**The handoff loses the evidence.** The parent receives a summary, not the material. It cannot tell a confident summary of nothing from a good one, and it has no way to check a figure. Whatever the subagent got wrong is now laundered into a clean sentence.\n\n**Errors compound the same way they always did.** Five subagents at 90% each give you a fully correct set about 59% of the time. Splitting the work did not change that arithmetic; it only made it harder to see.\n\n> Figure: A subagent is a fresh message list, not a colleague. What it genuinely buys: A separate context window: five suppliers researched costs the parent 2,000 tokens instead of 150,000; Real parallelism, so five independent lookups take the wall-clock time of one; A budget and a tool list belonging to that subtask alone. What it costs, every time: Tokens multiply; one published multi-agent system reported roughly fifteen times an ordinary chat interaction; The parent receives a summary and cannot check a figure inside it; Five workers at 90 per cent each are all correct about 59 per cent of the time; Workers cannot see each other's findings, so shared state breaks the split. Split when the reason is context isolation or parallelism, and write that reason down in one sentence. If you cannot, you have a workflow, and a workflow is cheaper, faster and easier to debug.\n\n## Three patterns worth knowing\n\n**Orchestrator and workers.** Parent decomposes, workers execute independently, parent assembles. Works when the subtasks do not depend on each other. Fails badly when they do, because workers cannot see each other's findings.\n\n**Pipeline.** Fixed stages, each with its own prompt: extract, then verify, then draft. This is a workflow rather than a multi-agent system, and being a workflow is why it is the most reliable of the three.\n\n**Critic.** A second call reviews the first. This helps when the critic has something the author did not: the test output, the schema, the original document, the validator's complaint. A critic with exactly the same information as the author mostly agrees with the author, at double the price.\n\n## When not to split\n\n- The subtasks share state or must happen in order.\n- A subtask can fail in a way the parent must handle precisely, rather than as a sentence.\n- The split mirrors your org chart. A researcher agent, a writer agent and an editor agent is a description of a newsroom, not a system design.\n\n## The rule\n\nSplit when the reason is context isolation or parallelism, and write that reason down. If you cannot state it in one sentence, you have a workflow, and a workflow will be cheaper, faster and easier to debug."},"check":{"question":"A team splits a document-processing agent into extractor, validator and summariser subagents, each running in sequence with the previous one's summary as input. Accuracy falls compared with the single agent. What best explains it?","options":[{"text":"The subagents lack shared memory, so they cannot coordinate their strategies.","correct":false,"tempting":"Reaches for coordination as the missing ingredient when the stages are sequential and, by design, do not need to coordinate."},{"text":"Three models are being sampled instead of one, so randomness has tripled.","correct":false,"tempting":"Blames sampling variance for a structural problem; the same three stages with temperature at zero would still lose the detail at each handoff."},{"text":"The stages are sequentially dependent and each handoff passes a summary instead of the source, so the validator and summariser check text rather than evidence and per-stage errors compound.","correct":true},{"text":"Subagents cannot use tools, so the validator has no way to verify anything.","correct":false,"tempting":"States a limitation that does not exist. A subagent is an ordinary loop and can hold any tools you give it."}]},"prev":{"slug":"agent-memory","title":"Memory is a directory, not a mystery","url":"https://addaly.com/learn/agents-and-automation/agent-memory"},"next":{"slug":"where-the-data-lives","title":"Files, inboxes, sheets and APIs","url":"https://addaly.com/learn/agents-and-automation/where-the-data-lives"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'When a second agent helps, and when it just costs more', in Agents and Automation. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/agents-and-automation/multi-agent-or-not (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"When a second agent helps, and when it just costs more","courseTitle":"Agents and Automation","module":{"index":2,"slug":"how-the-loop-runs","title":"The machinery: transcript, tools, loop"},"answer":null,"keyIdea":"A subagent buys you a separate context window and nothing else; pay for it only when isolation or genuine parallelism is the problem you have.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"left":{"title":"What it genuinely buys","items":["A separate context window: five suppliers researched costs the parent 2,000 tokens instead of 150,000","Real parallelism, so five independent lookups take the wall-clock time of one","A budget and a tool list belonging to that subtask alone"]},"type":"compare","right":{"title":"What it costs, every time","items":["Tokens multiply; one published multi-agent system reported roughly fifteen times an ordinary chat interaction","The parent receives a summary and cannot check a figure inside it","Five workers at 90 per cent each are all correct about 59 per cent of the time","Workers cannot see each other’s findings, so shared state breaks the split"]},"title":"A subagent is a fresh message list, not a colleague","caption":"Split when the reason is context isolation or parallelism, and write that reason down in one sentence. If you cannot, you have a workflow, and a workflow is cheaper, faster and easier to debug."},"alt":"Figure: A subagent is a fresh message list, not a colleague. What it genuinely buys: A separate context window: five suppliers researched costs the parent 2,000 tokens instead of 150,000; Real parallelism, so five independent lookups take the wall-clock time of one; A budget and a tool list belonging to that subtask alone. What it costs, every time: Tokens multiply; one published multi-agent system reported roughly fifteen times an ordinary chat interaction; The parent receives a summary and cannot check a figure inside it; Five workers at 90 per cent each are all correct about 59 per cent of the time; Workers cannot see each other’s findings, so shared state breaks the split. Split when the reason is context isolation or parallelism, and write that reason down in one sentence. If you cannot, you have a workflow, and a workflow is cheaper, faster and easier to debug."}},"blocks":[{"t":"h2","id":"what-a-subagent-actually-is","runs":[{"text":"What a subagent actually is"}]},{"t":"p","runs":[{"text":"Strip the diagrams away. A subagent is a fresh message list, with its own tools, its own budget and its own loop, whose final message comes back to the parent as a tool result."}]},{"t":"p","runs":[{"text":"That is the entire mechanism. There is no new capability, no emergent teamwork, no division of expertise beyond the two prompts you wrote. A “team of five specialist agents” is five prompts and five while loops."}]},{"t":"p","runs":[{"text":"Saying it plainly matters, because the diagrams imply an organisation and the code is a function call."}]},{"t":"h2","id":"the-one-thing-it-buys","runs":[{"text":"The one thing it buys"}]},{"t":"p","runs":[{"text":"Context isolation.","b":true},{"text":" A parent needs five suppliers researched. Done in one context, that is five sets of search results (perhaps 150,000 tokens) dragged through every remaining step. Done as five subagents, each burns its 30,000 tokens in its own window and returns 400 words. The parent’s transcript grows by 2,000 tokens instead of 150,000."}]},{"t":"p","runs":[{"text":"That is real, it is large, and it is the main honest reason to split. Genuine parallelism is the second: five independent lookups can run at once, and wall-clock time drops accordingly."}]},{"t":"h2","id":"what-it-costs","runs":[{"text":"What it costs"}]},{"t":"p","runs":[{"text":"Tokens multiply.","b":true},{"text":" Each subagent re-sends its own growing transcript, and the parent still pays for orchestration. One published account of a multi-agent research system reported it using on the order of fifteen times the tokens of an ordinary chat interaction. Budget for a multiple, not a margin."}]},{"t":"p","runs":[{"text":"The handoff loses the evidence.","b":true},{"text":" The parent receives a summary, not the material. It cannot tell a confident summary of nothing from a good one, and it has no way to check a figure. Whatever the subagent got wrong is now laundered into a clean sentence."}]},{"t":"p","runs":[{"text":"Errors compound the same way they always did.","b":true},{"text":" Five subagents at 90% each give you a fully correct set about 59% of the time. Splitting the work did not change that arithmetic; it only made it harder to see."}]},{"t":"h2","id":"three-patterns-worth-knowing","runs":[{"text":"Three patterns worth knowing"}]},{"t":"p","runs":[{"text":"Orchestrator and workers.","b":true},{"text":" Parent decomposes, workers execute independently, parent assembles. Works when the subtasks do not depend on each other. Fails badly when they do, because workers cannot see each other’s findings."}]},{"t":"p","runs":[{"text":"Pipeline.","b":true},{"text":" Fixed stages, each with its own prompt: extract, then verify, then draft. This is a workflow rather than a multi-agent system, and being a workflow is why it is the most reliable of the three."}]},{"t":"p","runs":[{"text":"Critic.","b":true},{"text":" A second call reviews the first. This helps when the critic has something the author did not: the test output, the schema, the original document, the validator’s complaint. A critic with exactly the same information as the author mostly agrees with the author, at double the price."}]},{"t":"h2","id":"when-not-to-split","runs":[{"text":"When not to split"}]},{"t":"ul","items":[[{"text":"The subtasks share state or must happen in order."}],[{"text":"A subtask can fail in a way the parent must handle precisely, rather than as a sentence."}],[{"text":"The split mirrors your org chart. A researcher agent, a writer agent and an editor agent is a description of a newsroom, not a system design."}]]},{"t":"h2","id":"the-rule","runs":[{"text":"The rule"}]},{"t":"p","runs":[{"text":"Split when the reason is context isolation or parallelism, and write that reason down. If you cannot state it in one sentence, you have a workflow, and a workflow will be cheaper, faster and easier to debug."}]}],"check":{"question":[{"text":"A team splits a document-processing agent into extractor, validator and summariser subagents, each running in sequence with the previous one’s summary as input. Accuracy falls compared with the single agent. What best explains it?"}],"options":[{"text":[{"text":"The subagents lack shared memory, so they cannot coordinate their strategies."}],"correct":false,"tempting":[{"text":"Reaches for coordination as the missing ingredient when the stages are sequential and, by design, do not need to coordinate."}]},{"text":[{"text":"Three models are being sampled instead of one, so randomness has tripled."}],"correct":false,"tempting":[{"text":"Blames sampling variance for a structural problem; the same three stages with temperature at zero would still lose the detail at each handoff."}]},{"text":[{"text":"The stages are sequentially dependent and each handoff passes a summary instead of the source, so the validator and summariser check text rather than evidence and per-stage errors compound."}],"correct":true,"tempting":null},{"text":[{"text":"Subagents cannot use tools, so the validator has no way to verify anything."}],"correct":false,"tempting":[{"text":"States a limitation that does not exist. A subagent is an ordinary loop and can hold any tools you give it."}]}],"order":[3,1,0,2],"why":null},"prev":{"slug":"agent-memory","title":"Memory is a directory, not a mystery","minutes":8},"next":{"slug":"where-the-data-lives","title":"Files, inboxes, sheets and APIs","minutes":9}}}