Skip to main content
An agent is a long-running autonomous worker scoped to a slice of your inventory. You give it a goal in plain English, set guardrails, and it runs on its own — triggered by new uploads, finished mapping jobs, mismatch reports, golden-dataset changes, or on a cron schedule. Every action it takes is recorded, reversible for 7 days, and capped by daily spend and hourly rate limits you control. Agents are available on Pro Max and Enterprise plans only.

​
When to use an agent

Always-on inventory hygiene

New rows uploaded every day? An agent re-runs mapping jobs automatically and files proposed matches for unmatched rows.

Mismatch triage at scale

Hundreds of mismatch reports backing up? An agent groups similar ones, drafts proposed resolutions, and queues them for a human reviewer.

Coverage monitoring

Want to know the moment supplier coverage drops below 95%? An agent watches, alerts, and posts a weekly digest to Slack.

Inventory in S3 or webhook

Drop CSVs into a bucket on a schedule and the agent pulls them, runs the job, and (optionally) pushes results back to a sink of your choosing.

​
When NOT to use an agent

  • One-off cleanup tasks — use the UI or a one-shot mapping job instead.
  • Bulk partner-hotel edits — agents cannot perform these in v1.
  • Auto-linking unmatched rows without human review — agents propose, humans approve.
  • Confidence-threshold-based auto-approval — see Rules and decisions for the orchestration-first framing in v1.

​
How it works (one cycle)

1

Trigger fires

An inventory upload completes, a cron tick lands, a mismatch is reported, or a user asks the agent a question in chat.
2

Runtime loads context

The agent runtime pulls the agent’s goal, current rule set, recent memory notes, open pending items, and the trigger payload.
3

LLM decides

Claude (or your BYO model) plans the cycle and calls tools — read-only lookups, mutating actions, and state writes.
4

Every step is persisted

Audit log entries, undo snapshots (for mutating tools), token + cost ledger updates, and notifications fire as configured.
5

Summary lands

The cycle ends with a 1-3 sentence summary written by the agent. Visible in the UI run timeline.

​
Trust by design

7-day undo

Every mutating action is reversible for 7 days from the workspace Undo Center.

Hard cost caps

Daily spend ceiling per agent — never blown by a runaway prompt.

Human-in-the-loop

Match proposals and mismatch resolutions land as pending items, not auto-applied changes.

​
What’s next

​
v1 roadmap (not yet shipped)

The following items are explicitly out of scope for v1 and will land in later releases. They are documented here so you know what to expect — and what not to expect — at launch.
  • Generated API reference for agent endpoints. REST endpoints exist (/api/v1/agents/*) but are not yet rendered in the API Reference tab. Use this section’s reference pages until generation lands.
  • Agent template library. v1 ships with 3 canned starter templates on the create flow. A broader marketplace is a v2 surface.
  • Multi-agent recipes / handoffs. Agents run independently in v1. Coordinated multi-agent workflows are deferred.
  • Customer-defined trigger DSL. Triggers are limited to the built-in set (inventory upload, mapping job finished, mismatch report, golden-dataset change, cron).
  • Confidence-threshold auto-approval rules. v1 is orchestration + triage only; per-row confidence judgement is not configurable yet.
  • SFTP and GCS data sources/sinks. S3 and HTTPS webhook only in v1.