Transcript
Agent Manager
This is the Agent Manager. An agent is a helper for routine work: instructions in plain words, a set of tools and a budget. Keldmoor's accounts payable team runs two: one files supplier mail, the other checks every invoice. Five tabs: Overview, Agents, Studio, Orchestration and Activity. No code and no IT project – the people who know the work set it up, and every run is budgeted, audited and under their control.
Your first agent
With no agents yet, the Agents tab offers a guided first agent – or a blueprint: a ready-made agent for a common job. Start, and pick what it should do. Keldmoor's choice: read mail and file documents. Choose the mailbox it watches – the AP mailbox, already set up in Connection Hub. Then the settings: the folder, the senders it may read – Keldmoor's suppliers – and how often to check: every five minutes. The test runs the agent on a sample message, as a plan – nothing is written. Once its test cases pass, it goes live – and files supplier documents into DocReveal on its own.
Instructions
For invoices, the team opens the agent wizard: New agent – six steps from a blueprint to a published agent. The Invoice triage blueprint reacts to every validated document: it approves clean invoices and escalates failed or suspicious ones. Give it a name and a description – the address is made from the name. The instructions are its job description: read the validation result, let a passed invoice through, hold a failed one with a label – and ask a person what to do next. Every save is a new version, and the Diff shows exactly what changed.
Tools
Capabilities are the tools it may use – here, ten: read the document, its validation result and mapped fields, ask a human, remember, label. Add more from DocReveal, your connectors and built-in tools – or hand work to other agents and flows. Tools that only read always run. Tools that change data are marked sensitive, and their group's policy decides: run automatically, ask for confirmation, or wait for a person's approval.
Triggers
Triggers start the agent on its own. This one reacts to an event: every document that finished validation. Others run on a schedule, with presets and your time zone, watch a mailbox, or start on demand. One switch exposes the agent to external clients. Off, it stays internal. The summary checks the definition against your tariff – and publishing waits until the agent passes its eval set.
Studio
In the Studio, the run input plays the event the agent will receive: a second copy of an invoice already in DocReveal. Chat with the draft and watch it work: the validation result, the mapped fields, the document profile – every tool call is listed. One rule failed: the invoice number must be unique per supplier. It won't approve a duplicate – it asks you instead. Answer, and it lays out its plan: hold the copy with a label, remember the case. Nothing is written – this is a test. A test run does the same unattended, and the trace records every step: model calls, policies, tool calls – and the wait for your answer.
Publish
Wednesday, the team sharpens the instructions. The draft changed, so publishing is blocked until the eval set passes again. The eval set runs the agent against its test cases – from the blueprint, and your own if you add them. Hard checks are exact: the right tools, the right result. Soft checks are graded by a model against a rubric. GO: both cases pass, and the verdict unlocks publishing – for exactly this draft. Publish… shows the summary once more – one click, and version three goes live.
Activity
Thursday afternoon, a Stavlund invoice: its line items add up to 8,240 dollars – the printed total says 8,420. The agent holds it and asks. The question waits in Activity, under Waiting for approval. Open the run: the question comes with ready answers. Pick one – and it continues where it stopped. Finished: escalated and labeled Do not pay – disputed. Your decision is on record. Agent runs lists every run of every agent. Recent activity is the audit of what they did.
Memory and teams
The run also proposed a note for next time: Stavlund invoices that don't add up are held, and a corrected one is requested. Nothing enters the agent's memory until a person confirms it – then it is part of every future run. Run journals record what each run did: the tools, the decision, the question. Orchestration maps how agents and flows work together: the validated event starts the triage agent; a flow sends approved invoices to the ERP. Click an agent to add a delegation – work handed to another agent. Chains stop after two hops.
Stay in control
The Overview shows your tariff's limits and what you have used: runs, tokens, the autonomy ceiling. The brain card is a short company profile every agent receives – generated from your documents, or written by you. Pause an agent with one click; resume it the same way. Export an agent and import it as a new draft – secrets never leave, and triggers arrive switched off. In Settings, Agent Manager holds the defaults: autonomy, budgets, memory and notifications. Building agents and answering their questions takes the administrator role – the rest of the team can follow every run. Two agents, built by the accounts payable team themselves – no code, and every run under their control.