[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F4MBE6PAH3H1TR7XSFCMPZ":3,"skill-same:01M7F4MBE6PAH3H1TR7XSFCMPZ":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":6,"ownerHtmlUrl":30,"ownerBlogUrl":6,"ownerDescription":31,"updatedAt":32,"install":33,"resources":45},"01M7F4MBE6PAH3H1TR7XSFCMPZ","skill-agent-topology",null,"Audit whether a multi-agent setup earns its coordination cost — use before adding an agent, or when a workflow feels slow or agents agree without adding signal","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fskill-agent-topology-claude-octopus\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: skill-agent-topology\ndisable-model-invocation: true\ndescription: \"Audit whether a multi-agent setup earns its coordination cost — use before adding an agent, or when a workflow feels slow or agents agree without adding signal\"\n---\n\n# Agent Topology Audit\n\nMost advice about multi-agent systems is about how to add agents. This is about\nwhether to. It audits a setup you already have, counts what each boundary\nbetween agents costs, and compares that against what the boundary buys. Removing\nan agent is a valid, and often the correct, result.\n\nThe framing comes from Liu, Canhui (2026), *The Organizational Behavior of\nAgentic AI* ([arXiv:2606.30986](https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30986)), which models\ncoordination overhead as **contextual transaction cost** — the cost of making\ntask context usable across an agent boundary.\n\n## When To Use\n\n- Before adding another agent, seat, or phase to a workflow that already works.\n- When a workflow is slow and it is not obvious which part is earning its time.\n- When agents keep agreeing. Agreement that costs three dispatches and produces\n  what one would have produced is overhead wearing the costume of consensus.\n- When a handoff keeps losing something and the fix keeps being \"add more\n  context to the prompt\".\n- After a workflow produced a bad result and you want to know whether the\n  topology or the models were at fault.\n\n## When Not To Use\n\n- To pick a workflow for a new task. That is `\u002Focto:auto`, which already routes\n  by intent, or `skill-decision-support` for a general option comparison.\n- To decide whether to delegate a task to agents at all. That is the allocation\n  step in `skill-intent-contract`.\n- To choose between providers or models. See `skills\u002Fblocks\u002Ffrontier-model-routing.md`.\n- For a single-agent task. There are no boundaries to count.\n\n## Inputs\n\n- The workflow or setup under audit: which agents or seats, in what order, with\n  what passing between them.\n- What each agent receives and what it returns. Prompt and output shape matter\n  more than model identity here.\n- Optionally, a transcript or run directory, which turns estimates into\n  observations.\n\nIf the setup is only described rather than run, say so in the output. An audit\nof a described topology is a prediction; an audit of a transcript is a\nmeasurement.\n\n## Workflow\n\n### 1. Draw the boundaries\n\nList every point where context crosses from one agent to another. Include the\nentry boundary (human to first agent) and the exit boundary (last agent to\nhuman) — they cost too, and the exit boundary is where synthesis quality is\nusually won or lost.\n\nCount them. The number of boundaries, not the number of agents, is what drives\ncoordination cost. Three agents in a star cost fewer crossings than three in a\nchain.\n\n### 2. Name what is lost at each boundary\n\nFor each crossing, work through these and record only the ones that actually\napply. Naming a cost that is not present is as unhelpful as missing one:\n\n- **Token and latency burden** — what it costs to restate context.\n- **Handoff** — what the receiving agent needs that the sending agent held but\n  did not pass.\n- **Compression loss** — what got summarised away. Free-text summary between\n  agents is the usual culprit.\n- **Semantic drift** — where the receiver's reading of a term differs from the\n  sender's.\n- **Verification burden** — work spent checking the other agent rather than\n  doing the task.\n- **Governance** — approvals, gates, and waiting.\n\n### 3. Name what the boundary buys\n\nA boundary is earned only by a gain that a single agent could not produce:\n\n- **Specialisation** — genuinely different capability, not a different label on\n  the same model.\n- **Parallelism** — real wall-clock reduction on independent work.\n- **Cross-vendor diversity** — different training data and different blind\n  spots. Note that same-family agreement is not this; see\n  `skills\u002Fblocks\u002Ffrontier-model-routing.md`.\n- **Adversarial review** — a seat whose job is to disagree, where disagreement\n  is the product.\n\n### 4. Compare against the single-expert null\n\nThe baseline is always one capable agent doing the whole task. The cited\nresearch found human-imitation topologies — pipelines, manager hierarchies, and\ncommittees deliberating in free text — measuring *below* that baseline, while\nagent-native forms built around shared memory measured above it. The single\nexpert stays competitive precisely because it pays no internal transaction cost.\n\nSo the **burden of proof falls on the boundary**. Absent a gain term that a\nsingle agent could not deliver, the recommendation is to collapse.\n\nTreat this as a directional prior, not proof. It is one simulation study plus\nmodel traces, and it is the source of the framing rather than a measurement of\nyour setup. Effect sizes from that paper are deliberately not reproduced here:\nthey describe the study's conditions, not yours.\n\n**One caveat that changes the reading, and must not be skipped.** What the study\npenalised was committee deliberation in free text with no independent evidence —\nagents talking to each other about the same information. Providers that bring\ngenuinely independent evidence, different models with different training data\nand real web search, are not that committee. `\u002Focto:debate` and `\u002Focto:council`\nare therefore better positioned than the studied form. The problem those results\nidentify is the handoff, not the panel.\n\n### 5. Reuse the overlap gate that already exists\n\nDo not invent a new \"is this agent adding anything\" metric. The council roster\nalready has one: `council_persona_overlap_score` in `scripts\u002Flib\u002Fcouncil.sh`\ncomputes a Jaccard index over persona capability tokens, and the roster builder\ndrops a candidate above `OCTOPUS_COUNCIL_DEDUP_THRESHOLD` (default 0.65).\n\nApply the same idea one level down. Two agents whose *inputs* overlap that\nheavily are usually one agent with two prompts.\n\n## Provider Or Data Priority\n\n1. A real transcript or run directory, if one exists.\n2. The workflow definition, for boundaries not exercised in that run.\n3. The user's description, flagged as unverified.\n\nPrefer counting observed crossings over reasoning about intended ones. Workflows\nroutinely skip or repeat boundaries at runtime.\n\n## Stop Or Checkpoint Rules\n\n- Stop and ask if the setup cannot be enumerated. An audit of a topology you had\n  to guess at is not evidence.\n- Stop before recommending removal of a boundary that exists for a safety,\n  approval, or compliance reason. Coordination cost is not the only axis, and\n  this diagnostic does not price the others.\n- If every boundary is earned, say so and stop. A clean audit is a real result;\n  manufacturing a finding to look useful is worse than none.\n\n## Output Contract\n\nReport in this order:\n\n1. **Verdict** — one of: collapse to a single agent; remove specific boundaries;\n   keep as is; restructure from chain to shared context.\n2. **Boundary table** — one row per crossing: what crosses, dominant cost term,\n   gain term claimed, and whether the gain is real.\n3. **What the single expert would have produced** — the null, stated concretely\n   enough to compare against.\n4. **For each boundary kept, what carries context across it** — the specific\n   artifact, file, or state that survives the crossing. A boundary kept without\n   naming this is a boundary that will keep losing information.\n5. **Confidence and basis** — measured from a transcript, or predicted from a\n   description.\n\n## Verification\n\n- Every boundary in the table maps to a real handoff in the setup, nameable in\n  the workflow definition or transcript.\n- Every claimed gain names the specific thing a single agent could not have\n  produced. \"Diversity\" alone does not qualify; different training data or a\n  different evidence source does.\n- Any cost term listed is one you can point at a concrete instance of.\n- The verdict follows from the table. If the table shows no earned boundary and\n  the verdict is \"keep as is\", one of the two is wrong.\n","# Agent Topology Audit\n\nMost advice about multi-agent systems is about how to add agents. This is about\nwhether to. It audits a setup you already have, counts what each boundary\nbetween agents costs, and compares that against what the boundary buys. Removing\nan agent is a valid, and often the correct, result.\n\nThe framing comes from Liu, Canhui (2026), *The Organizational Behavior of\nAgentic AI* ([arXiv:2606.30986](https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30986)), which models\ncoordination overhead as **contextual transaction cost** — the cost of making\ntask context usable across an agent boundary.\n\n## When To Use\n\n- Before adding another agent, seat, or phase to a workflow that already works.\n- When a workflow is slow and it is not obvious which part is earning its time.\n- When agents keep agreeing. Agreement that costs three dispatches and produces\n  what one would have produced is overhead wearing the costume of consensus.\n- When a handoff keeps losing something and the fix keeps being \"add more\n  context to the prompt\".\n- After a workflow produced a bad result and you want to know whether the\n  topology or the models were at fault.\n\n## When Not To Use\n\n- To pick a workflow for a new task. That is `\u002Focto:auto`, which already routes\n  by intent, or `skill-decision-support` for a general option comparison.\n- To decide whether to delegate a task to agents at all. That is the allocation\n  step in `skill-intent-contract`.\n- To choose between providers or models. See `skills\u002Fblocks\u002Ffrontier-model-routing.md`.\n- For a single-agent task. There are no boundaries to count.\n\n## Inputs\n\n- The workflow or setup under audit: which agents or seats, in what order, with\n  what passing between them.\n- What each agent receives and what it returns. Prompt and output shape matter\n  more than model identity here.\n- Optionally, a transcript or run directory, which turns estimates into\n  observations.\n\nIf the setup is only described rather than run, say so in the output. An audit\nof a described topology is a prediction; an audit of a transcript is a\nmeasurement.\n\n## Workflow\n\n### 1. Draw the boundaries\n\nList every point where context crosses from one agent to another. Include the\nentry boundary (human to first agent) and the exit boundary (last agent to\nhuman) — they cost too, and the exit boundary is where synthesis quality is\nusually won or lost.\n\nCount them. The number of boundaries, not the number of agents, is what drives\ncoordination cost. Three agents in a star cost fewer crossings than three in a\nchain.\n\n### 2. Name what is lost at each boundary\n\nFor each crossing, work through these and record only the ones that actually\napply. Naming a cost that is not present is as unhelpful as missing one:\n\n- **Token and latency burden** — what it costs to restate context.\n- **Handoff** — what the receiving agent needs that the sending agent held but\n  did not pass.\n- **Compression loss** — what got summarised away. Free-text summary between\n  agents is the usual culprit.\n- **Semantic drift** — where the receiver's reading of a term differs from the\n  sender's.\n- **Verification burden** — work spent checking the other agent rather than\n  doing the task.\n- **Governance** — approvals, gates, and waiting.\n\n### 3. Name what the boundary buys\n\nA boundary is earned only by a gain that a single agent could not produce:\n\n- **Specialisation** — genuinely different capability, not a different label on\n  the same model.\n- **Parallelism** — real wall-clock reduction on independent work.\n- **Cross-vendor diversity** — different training data and different blind\n  spots. Note that same-family agreement is not this; see\n  `skills\u002Fblocks\u002Ffrontier-model-routing.md`.\n- **Adversarial review** — a seat whose job is to disagree, where disagreement\n  is the product.\n\n### 4. Compare against the single-expert null\n\nThe baseline is always one capable agent doing the whole task. The cited\nresearch found human-imitation topologies — pipelines, manager hierarchies, and\ncommittees deliberating in free text — measuring *below* that baseline, while\nagent-native forms built around shared memory measured above it. The single\nexpert stays competitive precisely because it pays no internal transaction cost.\n\nSo the **burden of proof falls on the boundary**. Absent a gain term that a\nsingle agent could not deliver, the recommendation is to collapse.\n\nTreat this as a directional prior, not proof. It is one simulation study plus\nmodel traces, and it is the source of the framing rather than a measurement of\nyour setup. Effect sizes from that paper are deliberately not reproduced here:\nthey describe the study's conditions, not yours.\n\n**One caveat that changes the reading, and must not be skipped.** What the study\npenalised was committee deliberation in free text with no independent evidence —\nagents talking to each other about the same information. Providers that bring\ngenuinely independent evidence, different models with different training data\nand real web search, are not that committee. `\u002Focto:debate` and `\u002Focto:council`\nare therefore better positioned than the studied form. The problem those results\nidentify is the handoff, not the panel.\n\n### 5. Reuse the overlap gate that already exists\n\nDo not invent a new \"is this agent adding anything\" metric. The council roster\nalready has one: `council_persona_overlap_score` in `scripts\u002Flib\u002Fcouncil.sh`\ncomputes a Jaccard index over persona capability tokens, and the roster builder\ndrops a candidate above `OCTOPUS_COUNCIL_DEDUP_THRESHOLD` (default 0.65).\n\nApply the same idea one level down. Two agents whose *inputs* overlap that\nheavily are usually one agent with two prompts.\n\n## Provider Or Data Priority\n\n1. A real transcript or run directory, if one exists.\n2. The workflow definition, for boundaries not exercised in that run.\n3. The user's description, flagged as unverified.\n\nPrefer counting observed crossings over reasoning about intended ones. Workflows\nroutinely skip or repeat boundaries at runtime.\n\n## Stop Or Checkpoint Rules\n\n- Stop and ask if the setup cannot be enumerated. An audit of a topology you had\n  to guess at is not evidence.\n- Stop before recommending removal of a boundary that exists for a safety,\n  approval, or compliance reason. Coordination cost is not the only axis, and\n  this diagnostic does not price the others.\n- If every boundary is earned, say so and stop. A clean audit is a real result;\n  manufacturing a finding to look useful is worse than none.\n\n## Output Contract\n\nReport in this order:\n\n1. **Verdict** — one of: collapse to a single agent; remove specific boundaries;\n   keep as is; restructure from chain to shared context.\n2. **Boundary table** — one row per crossing: what crosses, dominant cost term,\n   gain term claimed, and whether the gain is real.\n3. **What the single expert would have produced** — the null, stated concretely\n   enough to compare against.\n4. **For each boundary kept, what carries context across it** — the specific\n   artifact, file, or state that survives the crossing. A boundary kept without\n   naming this is a boundary that will keep losing information.\n5. **Confidence and basis** — measured from a transcript, or predicted from a\n   description.\n\n## Verification\n\n- Every boundary in the table maps to a real handoff in the setup, nameable in\n  the workflow definition or transcript.\n- Every claimed gain names the specific thing a single agent could not have\n  produced. \"Diversity\" alone does not qualify; different training data or a\n  different evidence source does.\n- Any cost term listed is one you can point at a concrete instance of.\n- The verdict follows from the table. If the table shows no earned boundary and\n  the verdict is \"keep as is\", one of the two is wrong.\n",{"name":5,"description":7,"disable-model-invocation":24},true,"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","05-技能与代理开发\u002Fskill-agent-topology-claude-octopus",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":24},[35,38,41],{"tool":36,"command":37},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill skill-agent-topology",{"tool":39,"command":40},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill skill-agent-topology",{"tool":42,"command":43},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill skill-agent-topology","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fskill-agent-topology-claude-octopus\u002FSKILL.md",[],[47,54,59,66,73,80],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":19},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],{"skillNo":55,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":19},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]