docs/demo-script.md

# A five-minute TradeOpsSim demo

“Suppose GTS wanted to build an AI agent to monitor trade reconciliation, investigate mismatches, and flag cases that need review. AgentEnv provides a framework for building a test version of that workflow and evaluating the agent. Engineers define the records and permitted tools; operations specialists define a correct resolution or escalation. The framework runs the cases and records the actions and results.”

1. Open **The missed fill** and reset the episode. Show the 1,000-unit buy order: 400 filled, +400 booked, cancellation still pending.
2. Click **Run model assistant** on an authenticated host. The model receives only the business tools and operating runbook. On a host without model access, use **Guided workflow** and explicitly call it the rules baseline.
3. Open **Execution evidence**. Recovery reveals the delayed 200-unit execution, a second delivery of the same execution, and the cancellation acknowledgment.
4. Show the final state: 600 filled, 400 cancelled, zero working, and +600 booked. The correction is +200. The old 400-unit booking remains intact.
5. Show **Outcome checks** and **Action history**. The verifier checks records and evidence, rather than trusting the assistant's explanation.
6. Select **The unanswered cancel**, reset, and run the assistant. It can book the supported fill, but must keep the cancellation unconfirmed and open an evidence request. Correct escalation earns a passing assessment with an unresolved case status.
7. Optionally show **Same ID, different scope** or **Already reconciled** to demonstrate account isolation and the value of making no unnecessary change.

“GTS could start with a few sanitized, reviewed reconciliation incidents and turn them into repeatable tests. A later internal trial could give the agent copied records and have an operator review its proposed repairs. GTS would build that integration; AgentEnv would remain the place to rerun known cases after changing the prompt, model, or agent. Another workflow would use its own records, tools, and checks.”

The [internal workflow guide](agentenv-workflow.md) provides concrete responsibilities and commands for running the supplied baseline and model bundles.

All cases are synthetic. These authored scenarios demonstrate integration and measurable behavior; they are not a held-out performance benchmark or a live trading system.