{
  "edition": {
    "edition": "2026-09-06-draft-01",
    "language": "en",
    "publicationState": "published",
    "sourceRevision": "a7e788eb4b151fd92c1f58153e015773e3a829af",
    "projectCount": 67,
    "documentCount": 94,
    "skillCount": 10,
    "archivePath": "library.zip",
    "archiveBytes": 181812,
    "releasedAt": "2026-09-08T05:03:11.129+00:00"
  },
  "access": {
    "released": true,
    "rehearsal": false
  },
  "project": "agent-capabilities/model-consultation.md",
  "markdown": "https://legion.tw/api/projects/packet.md/agent-capabilities/model-consultation.md?edition=2026-09-06-draft-01",
  "documents": [
    {
      "path": "agent-capabilities/README.md",
      "text": "# Tools, skills, and plugins for the agent\n\n[All project types](../README.md) · [Shared launch](../shared/launch.md)\n\nExtend the coding agent's own working environment. The [shared skill package](../shared/skills.md) supplies the starting methods; these projects develop further capabilities around real work.\n\n## Projects to build\n\n- [Turn a working routine into a skill](routine-to-skill.md)\n- [Give the agent tools for an application](application-tools.md)\n- [Let the agent consult another model's abilities](model-consultation.md)\n- [Give the agent a continuing project notebook](project-notebook.md)\n- [Give the agent an ongoing responsibility](ongoing-responsibility.md)\n- [Package an agent capability for someone else](share-agent-capability.md)\n\nThese are instructions for projects you can build with your coding agent.\n"
    },
    {
      "path": "agent-capabilities/model-consultation.md",
      "text": "# Let the agent consult another model's abilities\n\nGive the person's coding agent a convenient way to consult another model when its particular abilities can improve the work. Build it extremely well. An agent editing audio should be able to obtain a real listening judgment; one preparing a report should be able to ask a strong visual model about the actual photographs. Follow the [shared launch](../shared/launch.md).\n\nChoose one recurring need in a project the person already uses. Complete that consultation inside the real workflow, keeping the existing agent responsible for the overall task.\n\n## Choose for the work\n\nDiscover which models and supported access routes fit the chosen job. Compare them using relevant material and the qualities the person cares about: understanding speech in a noisy room, recognizing a detail in an image, or writing naturally for a particular audience.\n\nConsider the full working route, including what inputs reach the model, which tools it has, waiting time, and cost. A model's name alone does not describe the experience the application will deliver.\n\nReuse suitable access the person already has. Explain any additional account or API key needed and what is billed through it. Keep credentials in their appropriate local configuration and make changing the consulted model practical as the person's options develop.\n\n## Send a colleague enough to reason with\n\nBuild a callable tool or suitable native integration through which the main agent can pose the question and supply the relevant material. Send actual images or audio when perception matters, along with enough surrounding context to interpret them.\n\nFrame the request as work for an intelligent colleague. Explain the purpose, audience, and judgment needed, leaving room for observations the main agent missed. Describe a useful outcome in positive terms.\n\nFor example, send the original recording and proposed edit with a request to judge whether the cut preserves the speaker's meaning and sounds natural. The consulting model should be able to identify the relevant moments, explain a problem, or conclude that the edit works well.\n\nKeep requests focused on the contribution being sought. When an independent reading matters, give the colleague the source and question before the main agent's preferred explanation. Choose context deliberately according to the kind of judgment needed.\n\n## Bring the contribution back into the work\n\nReturn the response and any useful referenced material in a form the main agent can use. Keep track of which model, inputs, and version of the work the consultation concerned. Make failed or incomplete calls recognizable so an absent answer remains distinct from a judgment.\n\nHave the main agent consider the colleague's reasoning, reconcile it with the actual task, and decide what to do. It can request a focused clarification when useful. Preserve sound work when the consultation supplies no reason to change it.\n\nMake the result visible in the person's ordinary workflow: an auditionable audio revision, a better image selection, or a clear explanation of what the additional model noticed. Keep the original material and useful prior versions available.\n\n## Deliver a useful consultation\n\nInstall the integration in the agent's current environment and make its purpose discoverable in relevant project guidance or a skill. Exercise it on the chosen task with real supported inputs, then inspect what the colleague actually received and returned.\n\nJudge the improvement through the work the person uses. Explain what left their machine, which capability the second model contributed, and how they can control usage or disable the connection.\n\nTeach that model choice can happen within a workflow. They can keep an agent they like working with and give it access to a colleague with different strengths.\n"
    }
  ]
}