AlexSchwartzman/TradeOpsSim-AgentEnv

TradeOpsSim

AgentEnv is a framework for building software environments and evaluating agents in them. TradeOpsSim is one example: engineers supply a trading workflow, its tools and business checks; the framework deploys the components, runs the cases and retains the results.

An original synthetic trading-operations environment: reconcile a delayed execution, a repeated delivery, and an interrupted cancellation without inventing missing evidence. Thirteen fixed cases cover clean records, partial recovery, account scope, rejected cancellations, completed orders, sell-side signs, and incomplete sources. No market data or production system is connected.

The rules baseline and AI agents use the same nine tools. The evaluator checks immutable source evidence separately. A correctly escalated unresolved case can pass; a green task status alone does not establish success. Final model tasks also require a completed, error-free agent response.

All native runs in the baseline and host export store are retained, including development attempts and infrastructure failures. Separate API contract tests and the video's UI runs are documented in the validation notes and video source package. The task_version field distinguishes earlier task graphs from the final combined reward graph. Runs that fail before a model prompt have no model trajectory and are excluded from model-quality denominators. These thirteen authored fixtures are a demonstration, not a held-out benchmark or evidence of production performance.

The portable baseline and API-agent bundles include their complete Docker build contexts. The host-CLI bundle records the actual model trials and requires the source runner.

Watch and try it

Watch the video and inspect recorded runs. The Space shows existing synthetic results. Download the environment to run your own agent.

Watch the narrated AgentEnv walkthrough. Suppose GTS wanted an agent to monitor trade reconciliation, investigate mismatches, and flag cases for review. The film shows how AgentEnv could help the team build a test environment and evaluate that agent. Two recorded synthetic reconciliation runs illustrate the checks; the GTS pilot and internal trial are proposals. The film combines actual application captures with explanatory diagrams, with captions and a transcript.

Using AgentEnv internally describes the records, tools and outcome checks a team supplies, the test execution AgentEnv manages, and concrete reproduction commands.

Download the application source for the operator UI, MCP server, all thirteen cases, tests, and AgentEnv publishing tools. The editable video package contains the film's local assets, original captures, run evidence, provenance, and pinned HyperFrames source.

To run the baseline, download this repository and use Python 3.11.4+ with Docker running:

python -m pip install 'agentenv-framework==0.9.1324'
agent-env run ./bundles/tradeops-baseline --task late_fill_duplicate

The optional HF plugin also supports downloading and running the bundle directly:

python -m pip install 'agentenv-hf @ git+https://github.com/earakely-scale/agentenv-hf-plugin@c116bf1b321fa33d671632b0ee1b77406dbe3b58'
agent-env hf run AlexSchwartzman/TradeOpsSim-AgentEnv --bundle tradeops-baseline --task late_fill_duplicate

The default bundle is tradeops-baseline; it requires no model key. Use tradeops-agent for a configured API model; its bundle README specifies runtime configuration. Reproduce the host CLI trials from the source archive, not with the standalone HF run command.

Recorded validation

Track	Published task rows	Published episode rows	Actual model trajectories
Rules baseline	13	14	0
Portable API agent	13	0	0
Host CLI model	13	28	27

The final task-version-2 model sweep passed all thirteen authored cases: nine reconciliations and four correct unresolved outcomes. The configured model identifier was claude-sonnet-5-5. The baseline passed the same thirteen cases. The baseline and host tracks each retain one earlier pre-prompt infrastructure failure with a null reward. The API adapter passed a full Docker/Task.run contract test using a scripted provider; that test is not an AI result. See validation for scope, negative controls, and reproduction commands.

Six Parquet configurations separate tasks from episodes for each track. raw/ preserves native run records and supported trajectories; bundles/ contains frozen executable artifacts. Task versions distinguish the early business-only score from the final score that also requires successful agent completion. A null reward denotes an ungraded infrastructure failure.

Transparent rules policy. No model or API key is required. This is a software baseline, not an AI result.

Portable A2A model bundle. Supply a compatible API endpoint, credential, and model at runtime. Its API contract is tested locally; the recorded real-model trials use the separately identified host CLI adapter.

Supplemental recorded host CLI trials. Reproduce with scripts/run_host_trials.py from the source distribution and an authenticated host CLI. This task bundle is not a standalone agent-env hf run target.