Document Agents

Document Agents are configurable AI generators for business documents — contracts, proposals, quotes, invoices, offer letters, and more — combined with an end-to-end e-signature workflow.

A Document Agent defines the type of document (the prompt, the tools, the model, the locked sections, the signer roles). A Document is one generated instance. Signing sessions track each recipient's path through the signing flow.

Key Concepts

• Document Agent — A reusable configuration: prompt, tools, model, hardcoded sections, signer roles, output formats, page setup
• Document — An instance generated from an agent for a specific customer / counterparty / project
• Hardcoded section — Verbatim "carbon-copy" text the renderer always inserts (e.g., governing law, disclaimers). The AI is told they exist and is instructed not to duplicate them
• Global section — Reusable headers, footers, clauses, or disclaimers shared across many document types
• Signing field — A signature, initial, date, text, or checkbox placeholder the AI places inline; the renderer resolves it to pixel coordinates on the PDF
• Signing session — Per-recipient token-bearing record; captures consent, identity attribution, signing action, and a full audit trail

Where to Find Them

• Document agents: Navigate to Agents > Document Agents to manage types and templates
• Generated documents: Navigate to Documents to see all generated instances and their signing status
• Global sections: Navigate to Documents > Global Sections to manage reusable headers/footers/clauses
• Template library: Navigate to Document Agents > Templates for the one-click preset gallery

Creating a Document Agent

The fastest path is to install a preset and edit it. To start from scratch:

1. Agents > Document Agents > New Document Agent
2. Name & description — what this type of document is for
3. Model — gemini-3.5-flash is the default and recommended for most documents
4. Instructions (system prompt) — tell the AI what sections to produce, in what order, what tone, and how to format the JSON output envelope (the form documents this)
5. Tools — same catalog as task agents (vector search, web search, save-update-user, send-email, send-sms, plus any custom tools your tenant has registered)
6. Output formats — pick any subset of PDF, DOCX, HTML
7. Hardcoded sections — paste in any verbatim legal boilerplate (governing law, severability, force majeure). Choose where in the layout each section appears
8. Global sections — pick existing headers/footers from your tenant's global section library
9. Default signers — define roles (e.g., "Disclosing Party", "Recipient") and order; real names/emails are filled in per document
10. Page setup — size, margins, font, font size, brand color
11. Attach to CRM records — which entity types this document type can be generated from (contacts, leads, companies, deals). Each CRM detail page only offers templates whose attachableTo list includes that entity type.

Save the agent. It now appears in your document agents list and can be used to generate new documents.

Generating from a CRM Record

Open any contact, lead, company, or deal. Scroll to the Documents card and click New document. The picker lists only the document types whose attachable-to list includes that entity. The generated document is tagged with attachedTo metadata and shows up in the same card on subsequent visits, plus on the global Documents list filtered by attachment.

Generating a Document

1. Documents > New Document (or click "Generate" from a document agent)
2. Fill in the input data the agent expects (party names, key dates, prices, etc.) — these are passed to the prompt as {placeholders}
3. Assign real names and emails to the signer roles
4. Click Generate with AI

A streaming progress panel shows:

• started — agent loaded, prompt assembled
• iteration — each round of model thinking
• tool_call / tool_result — when the AI invokes a tool
• parsed — model returned a valid JSON envelope
• section.delta — each section as it lands
• done — generation finished, document is in draft status

You can edit any AI-authored section in place. Hardcoded and global sections are locked (shown with a lock icon) so they can't be accidentally edited.

Placing Signing Fields

The AI auto-places signing fields based on the inline markers it emits during generation. You can refine or completely override that placement using the Place fields tab in the document editor:

• Each PDF page is rendered in-browser via PDF.js
• Pick the active signer from the toolbar (each signer gets its own color)
• Click an "Add" button (signature / initial / date / text / checkbox), then click anywhere on a page to drop the field
• Click and drag any existing field to move it; drag the bottom-right corner to resize
• Hover a field to reveal its delete button
• Click Save to persist; the fields are now stamped at exactly those coordinates when the document is signed

Field coordinates are stored in PDF point space (top-left origin), so they survive page-size and font changes consistently.

Field placement is locked once a document moves to out_for_signature — you can't reposition fields after invites are sent.

Approving and Sending for Signature

1. Review the draft. Click Approve when you're satisfied — the status moves to approved
2. Verify all signers have names and emails
3. Click Send for signature
   • Sequential mode — only the first signer (by order) gets the email; the next signer gets theirs as soon as the previous one finishes
   • Parallel mode — every signer gets the email at the same time
4. As each signer completes, the document status flows through out_for_signature → partially_signed → signed
5. Once all signers sign, the system stamps the final PDF with each signature image (or typed name), appends a Certificate of Completion, computes a SHA256 integrity hash, and emails a signed copy to every party

If anyone declines, the document moves to declined and no further invites are sent. You can void a document at any time before it's fully signed.

Audit Trail

Every meaningful event is recorded on the document with a timestamp, actor, IP address, and user agent:

• created — document was created
• generated — AI generation completed
• section_edited — a section was hand-edited
• approved — author approved the document
• sent_for_signature — sent to signers
• invite_sent — individual signer received an invite
• viewed_by_signer — signer opened their link
• signer_signed — signer completed their signature
• signer_declined — signer rejected the document
• finalized — final signed PDF stamped, hash recorded
• voided — document was voided by the author

This trail is embedded in the Certificate of Completion appended to the final PDF.

Common Use Cases

• Sales proposals — generate a tailored proposal per opportunity, send the same day
• NDA workflow — install the mutual NDA preset, fire it whenever a new partner conversation starts
• MSA + SOW pairing — sign the MSA once with a customer, then issue SOWs against it for each engagement
• Offer letters — generate from a recruiting workflow, route to hiring manager then candidate (sequential)
• Invoices — produce a clean PDF invoice from CRM line items, no signing required