the Skyflo log

An Agent That Carries Only the Tools It Needs

Skyflo 1.8.3 loads Mac, meeting and other specialist tools only when a mission needs them. Here is what that saved, what it cost at first, and how we fixed it before it shipped.

Read the full page:How Skyflo runs a mission, and learns from it
·5 min read·Karan Jagtiani·tool-searchcontextcomputer-usemeetings

Every request an AI agent sends to its model carries a description of every tool it might use. Skyflo's own agent knows how to read and edit files, run commands, search the web, work a Mac app window, read a meeting transcript and post into a call's chat. Until 1.8.3, a mission that only needed to read one file still sent the instructions for all 47 of those tools, on every request.

That is wasted input. It costs money, it takes time to read, and it gives the model more to sort through before it starts on what you asked.

We argued for tool search a long time ago: give the agent a short list of what exists and let it load a tool when it needs one. Skyflo 1.8.3 ships it in Skyflo's own agent, and we measured it before we let it out.

What changed

A mission now starts with 40 tools instead of 47. The ones most missions never touch, such as working a window in another Mac app or posting into a meeting chat, are listed by name with one line each. When the agent needs one, it asks for it and has it from its next step.

Skyflo also loads them up front when it already knows they will be needed:

  • A saved meeting. If the mission has a transcript from your call, it can read it from the first step.
  • A window it already has. If the mission already controls an app window, the Mac tools are there from the start.
  • Specialists in Orchestrate. A specialist that cannot search for tools gets the ones its role allows up front, so it is never stranded.

None of this widens what an agent may do. A tool loads only if the agent's role and the mission's access mode already allow it.

What it saved

We ran the same tasks on signed 1.8.3 builds with and without on-demand loading, with GPT-6.1 Terra at Low reasoning, three times each:

TaskRead two values from a file
Input before38,569 tokens
Input after29,855 tokens
Change23% less
TaskAnswer questions about a saved meeting
Input before63,480 tokens
Input after50,904 tokens
Change20% less
TaskFix a bug across two files and run the tests
Input before119,306 tokens
Input after105,545 tokens
Change12% less

Every answer was the same. The file values were exact, the meeting questions were all right, and the bug fix passed its tests while touching only the two files with the bug.

What it cost, and the fix

The first version was not a win everywhere. A task that started by working a Mac app in a new mission cost 27% more, counting input the provider had already cached at a tenth because it costs far less.

The reason was the cache. Model providers reuse the part of a request they have seen before, and changing the tool list starts that over. The agent loaded three of the five Mac tools, then picked up the rest after it was given the window, so the tool list changed twice. And when a specialist was given a window, the mission that sent it still had to search for the Mac tools itself.

We fixed both before shipping. A request for any Mac tool now loads the whole set at once, and a mission notices a window its specialist already holds. On the release build, the same Mac task costs 9% less than it did before on-demand loading. Across all 19 tasks in our test set, from file work to Mac apps to Orchestrate, the agent sends about 10% less to the model.

What we checked

Saving tokens is only worth it if nothing gets worse. Across 62 runs on signed builds, every task passed on every build. No agent edited a file outside its task or used a window it was not given. On-demand loading also passed on a second model family, Gemini 3.8 Flash.

Also in 1.8.3

  • Setup suggests your areas. Describe your company in the first-run tour and Skyflo proposes areas for Home.
  • Usage limits explained for every agent. When ChatGPT, Claude or Skyflo's own agent reaches a limit, the mission says which limit and when it resets.

The 1.8.3 release notes list everything in the release. Download Skyflo for Mac, or update from the app.

Loading AI Agent Tools on Demand, Measured