Chain mode: let the AI build your workbook end-to-end

Chain mode auto-applies every AI step, observes formula results, and continues building until your entire workbook plan is complete.

Chain mode removes the manual Apply step from the regular chat flow. Instead of reviewing and clicking Apply after each AI response, you send one prompt and GridOS loops — applying writes, reading formula results, and generating the next section — until the entire plan is done or you stop it. It’s designed for large multi-section builds where clicking through every step would be tedious.

When to use chain mode

Large financial models

Revenue, COGS, gross profit, OpEx, EBITDA, and a DCF terminal value all in one run — chain mode handles the full structure without you clicking Apply on each row.

Multi-section workbooks

Any model with more than three or four sections benefits from chain mode. The agent tracks which sections it has written and which remain.

Formula-dependent builds

Chain mode observes the computed values of formulas after each write, so it can use those values to inform subsequent sections — for example, using a computed revenue figure to seed the COGS row.

Template-like builds from prompts

Describe a full workbook in one prompt and let the agent build it end-to-end, just like opening a template but tailored exactly to your spec.

How chain mode works

1. Send your prompt

   Type your full build request in the chat composer and select Chain mode (instead of the default single-turn mode) before submitting.

   Build a 4-quarter operating model: revenue starting at 100 with 10% QoQ growth,
   COGS at 40% of revenue, gross profit, OpEx fixed at 20, and EBITDA
   
2. The agent declares a plan

   On the first turn, the agent emits a plan — a structured preview of every section it intends to fill, with target ranges and notes. This plan is shown to you before any cells are written. The agent also writes the first section only on this turn.

   Plan: Quarterly Operating Model
   ├── Header row       B2:F2   Metric, Q1–Q4
   ├── Revenue          B3:F3   100, 10% QoQ growth
   ├── COGS             B4:F4   =MULTIPLY(revenue, 0.4)
   ├── Gross Profit     B5:F5   =MINUS(revenue, COGS)
   ├── OpEx             B6:F6   Fixed 20 per quarter
   └── EBITDA           B7:F7   =MINUS(gross_profit, OpEx)
   
3. GridOS auto-applies each step

   After writing a section, GridOS applies it to the kernel immediately — no click required. It then reads back the computed values of any formula cells and feeds those observations into the agent’s next prompt.

4. The agent writes the next section

   Informed by the observed formula results (and any warnings), the agent writes the next section. This repeats until every section in the plan is complete.

5. Completion signal

   When all sections are done, the agent signals completion by returning values=[[""]] — an empty write. GridOS stops the loop and presents the final summary of all steps taken.

The plan preview

The plan is informational — it shows you the full structure the agent intends to build before cells are filled. It is emitted only on the first turn and suppressed on all subsequent turns.

Each section in the plan includes:

Field	Description
label	Human-readable name for the section (e.g. “Revenue”, “COGS”)
target	The cell range the agent will write (e.g. B3:F3)
notes	Brief description of the formula logic or values

The plan is a preview, not a commitment. If the agent detects that a section needs to be retried (because of a column-alignment warning), it rewrites that section before advancing — the plan section count may not match the number of loop iterations exactly.

Column alignment auto-retry

GridOS detects the most common formula error in financial models — referencing a label cell instead of the adjacent data cell — and triggers an automatic retry when it occurs.

After applying each section, GridOS checks whether any formula references a text-valued cell. If it does, the agent receives a detailed warning and is told to re-emit the same section with corrected cell references before moving to the next section.

*** COLUMN ALIGNMENT WARNING ***
Formula in C4 references non-numeric cell B3 (value: "Revenue").
Re-emit the COGS section with corrected formulas. Each formula in column X
must reference cells in column X, not the label column.

This retry does not count as a wasted iteration — sections are only counted as cleanly written when they produce no warnings.

Iteration limits

Chain mode runs for up to 10 iterations by default. Each application of a section (including retries) counts as one iteration.

If your model has more than 10 sections, chain mode will stop before completing it. Break the build into two prompts — run chain mode for the first half, then run it again for the second half.

You can reduce the limit by passing max_iterations in the API request if you want shorter runs during development.

Stopping chain mode

Chain mode stops automatically when any of the following happen:

• The agent returns an empty values=[[""]] completion signal
• The maximum number of iterations is reached
• The agent has no formula observations remaining and the plan is complete

You can also dismiss the chain at any step from the UI — the cells written so far are kept; only the remaining steps are cancelled.

Chain mode vs. regular chat

	Regular chat	Chain mode
Apply step	Manual — you click Apply after each response	Automatic — applied immediately after each step
Formula observation	Not fed back to the agent	Yes — computed values are included in the next prompt
Column alignment retry	You see the error and re-ask manually	Automatic — the agent retries the section before moving on
Plan tracking	Agent tracks progress informally	Agent tracks sections written vs. remaining, with explicit hints
Best for	Single writes, quick edits, questions	Large multi-section builds, full model generation

For large financial models, combine chain mode with templates: open a template to get the skeleton, then use chain mode to fill in the model with your specific assumptions in a single prompt.