QuantyDocs

Agents

An agent is a reusable AI persona: a name, a set of instructions, a model and optional connectors. Instead of re-explaining context in every conversation, your team builds an agent once — "Invoice Auditor", "Logistics Analyst" — and picks it from the chat header whenever that expertise is needed.

The Agents page with two custom agents

What an agent is

Each agent defines:

  • Name and description — how it appears in the picker (names are unique per organization).
  • Instructions — the standing brief injected into every conversation with this agent.
  • Model — which LLM the agent uses by default.
  • Color — the accent used for its avatar.
  • Connectors — which integrations the agent may lean on: gmail, google_sheets, outlook, slack, notion, hubspot.
  • Sample prompts — up to 6 starter prompts (300 characters each) shown when a conversation with the agent is empty.

The built-in Quanty agent is always available and marked Default; it is not stored as a record and cannot be edited or deleted.

Building agents

Open Agents in the sidebar. The builder page lists your organization's agents and lets anyone except Viewers create and edit them.

Generate with AI

The fastest path: describe the agent in plain language ("an analyst that reviews carrier invoices against our rate cards and flags overcharges") and let Quanty draft the whole definition — name, description, instructions, connectors and sample prompts. Review the draft, adjust, and save. Nothing is persisted until you save.

Manual editing

Every field is editable in the in-page editor. You can also generate sample prompts for an existing agent with one click; the three suggestions are stored on the agent immediately.

Activate, deactivate, delete

  • The is_active switch takes an agent out of rotation without deleting it — deactivated agents cannot be selected in chat.
  • Deleting is a soft delete: the agent disappears from lists, and its name is freed for reuse.

Using agents in chat

The chat header's agent switcher (a searchable picker) lists Quanty AI Agent plus your custom agents with their avatars. Selecting one applies its instructions to the conversation and shows its sample prompts in the empty state. See Chat.

Delegated agent tasks

For hard, multi-step questions, chat can hand work to an autonomous agent task — for example "find this supplier's parent company and their latest revenue". A task:

  • runs with an explicit budget (default $0.50, at most $2.00 per task), and stops when the budget is reached,
  • works step by step — searching, reading documents, computing — and records every step,
  • returns a value with confidence and reasoning, plus the actual cost.

Numeric work happens in a Python sandbox where possible. Results carry a badge: Computed in the Python sandbox for verified computation, or Not sandbox-verified — model-estimated value when the model estimated instead.

Budgets keep costs predictable

Every agent task has a hard budget and its spend is recorded against your organization's AI usage, like any other AI feature. Runs that hit the cap end with status "Budget reached" and keep partial results.

Agent activity

Delegated runs are visible live:

  • The Agent activity panel opens from the sheet toolbar and the chat header — a pulsing dot signals active runs.
  • Each run shows ordinal, timestamps and status: Queued, Running, Done, Failed, Stopped or Budget reached.
  • Open a run to see its full step trail; Stop cancels a run before its next step.
  • Runs are listed per sheet and organization-wide, so you can audit what agents did and what it cost.
  • Chat — where agents are picked and tasks are launched.
  • Billing & tokens — how agent spend is accounted.