Production agents, inspected

Agent Engineering Lab

A deterministic learning environment for tool boundaries, runtime schema validation, policy gates, and human approval — the parts of an AI agent the chatbot demos hide.

Credit order eligibility

The first workflow asks whether a customer can place a new order, then shows the messages, tools, policy checks, schema validation, and final recommendation.

Make the hidden loop visible

Agent Lab is a deterministic learning environment for the moving parts behind production AI agents. The model is simulated so the lab works without API keys; the loop, tools, schemas, policy, and approval gate are real.

Messages between user and model are visible
Tools have typed arguments and runtime-validated results
Policy is separate from the model and is consulted before every tool call
Write actions pass through a human-in-the-loop approval gate
Final recommendations are validated against a Zod schema before they are shown

Scenario

Mock data

Customers

ACME Corp: $38,000 / $50,000, watchlist
Globex: $22,000 / $120,000, active
Initech: $19,000 / $25,000, blocked

Open invoice signals

inv_acme_1017: $6,400, overdue, 42 days past due
inv_acme_1034: $11,800, open
inv_globex_2042: $9,000, open
inv_initech_3031: $7,200, overdue, 77 days past due

Interactive lab

A realistic ERP credit check where the agent must inspect customer status, invoices, policy, and ask for approval before creating a business record.

Can customer ACME place a new order for $20,000?

Idle— Pick a scenario and run the simulation

Model

Latency

0ms

Tool calls

Iterations

Tokens

simulated

Cost

$0.000000

simulated

Overview

Lesson summary and run controls

Architecture →

Pick a scenario, run the simulation, then explore the other tabs to see the same run from different angles: messages, schemas, tool calls, the loop itself, or the full trace.