Long jobs with an external AI agent
MindMax already runs these jobs with its own assistant, so why borrow your app's model? Three reasons, all practical: you may already pay for a model you would rather use, you may want a stronger model or parallel sub-agents for a big book, and this path spends no MindMax credits — the tokens go on the app's bill. MindMax still does the parts it is good at: planning the steps, cutting the material into self-contained packages, validating what comes back, and saving the result beside the source. The app's model does the thinking.
If you would rather hand the whole job to MindMax's own assistant, that is the same plan-pause-resume shape with a different entry point and a different bill. This page is the version where your app does the work.
Before you start
- An app connected with Collaborate access.
- The document or folder open in MindMax, and a clear idea of the result you want.
- A little patience: a whole book is many steps. A run keeps its progress, so you can stop and come back to it.
Three calls drive every run
Use your own model to run a study guide for the book I have open in MindMax, and save it where MindMax would save its own.
| Call | Who uses it | What it does |
|---|---|---|
start_pipeline | The app, once | Picks the method and its options; MindMax returns the plan and the first work package |
get_step | The app, repeatedly | Fetches one self-contained step: instructions, the input, the expected output shape |
submit_step | The app, repeatedly | Hands back the finished step; MindMax validates it, anchors quotes and saves it, then names the next steps |
Steps marked parallel can be given to sub-agents. A result that fails validation comes back with the
reason and can be resubmitted twice. When the run is done, open_in_mindmax puts the result in front
of you — the reading panel, the side-by-side translation view, the book's map, or the canvas.
If the app disconnects halfway, the run keeps what was already saved. Press Continue in MindMax, or ask the app to pick the run up again; completed steps are not redone. See common issues if a job appears stuck.
The seven methods you can borrow
| Method | What it produces | Option worth knowing |
|---|---|---|
reading_guide | A full annotation pass with margin notes and a closing write-up | lens (guide / review / critical / study), focus |
comment_answers | An answer in each waiting comment thread | — |
comment_insights | An insights note distilled from your highlights and comments | — |
doc_map_notes | Chapter and node notes written into a long book's map | — |
canvas_skill | One thinking method laid out on a canvas | skill, target |
translation | A whole-document translation with one terminology list | targetLanguage |
ocr | Recognised text for scanned pages, page by page | Page images come with each step |
The result lands in the same place a MindMax run would have put it, so the rest of your library can use it: quotes become searchable, chapter notes improve later answers, and a translated document opens in the same side-by-side view. The agent tool reference lists each method with its output shape.
What it costs, and what to compare
A full pass costs roughly one read of the material in the app's own context — real tokens, billed to the app, not to your MindMax balance. No MindMax model runs on this path. There is no fixed price to quote: a 20-page report and a 550-page book are different jobs, and the app should tell you before starting a long one.
The results are the same shape as MindMax's own runs. These pages show what each one looks like once it is in your folder, so you can decide which entry point suits the job:
- Translating a document and translate a book you need to read
- Recognising scanned PDFs
- Understand a paper outside your field and reading and reviewing with an agent
- Thinking on the canvas and canvas requests with an agent
- Run a long job with your agent — the same path as a worked use case
Common problems
The app wrote an answer in chat and no step was saved. It skipped submit_step. Ask it to
submit each step to MindMax rather than summarising the whole job at the end.
A step was returned with a reason. Read the reason, fix that step's output and resubmit. Two resubmissions are allowed before the run stops to ask you.
The run stopped when I closed the app. Progress is kept. Press Continue in MindMax or ask the app to resume; finished steps are not repeated.
The scan produced poor text. OCR quality depends on the scan. MindMax reports which pages came out badly instead of quoting them as clean text — look at the page images for those pages.
Next
- Controlling agent access — what is logged, what can be withdrawn, and how to pause an app.
- Working with an external AI agent — the full capability map.