← All work

ML Tooling / 2026

Agent evaluation environments

Testing how agents investigate and reason.

I built OpenEnv investigation environments with procedural cases, MCP tools, evolving state, and adversarial tests.

PythonOpenEnvMCPDocker

Hackathon projects

CONTRIBUTION & EVIDENCE
Run the workflow

Follow the investigation.

Walk through an investigation and see why evidence matters to evaluation. Click a step, connection, or trace to inspect it.

Execution trace / Click to inspect

Start the walkthrough, or inspect any step before running it.

Simplified from the project architecture. Runs locally; no live models, credentials, or external services.

Agent evaluation environments / Inspector

Workflow overview

Scenario

A procedural scenario establishes the initial evidence and environment state.

Why it matters

These environments evaluate actions in a simulated scenario. Evidence gathering and the evolving state matter alongside the final answer.

Synthetic payloads explain the boundaries. They are illustrative examples, not project API schemas or live production traces.

The engineering

Inside
the decisions.

Evaluate the investigation, including the route to an answer.

Evaluating an agent requires more than checking its final answer. The environment must test how it gathers evidence, sequences actions, and responds as the situation changes.

01

The design decision

Represent scenarios as evolving state with explicit tools and rewards. Test for shortcuts through adversarial cases.

02

The trade-off

A simulated reward is an approximation of good behaviour. Procedural variation and adversarial testing help expose weaknesses, but cannot establish real-world security or compliance performance.

03

Building a stateful investigation

Built a compliance investigation environment with MCP tools, scenario state, procedural generation, and a multi-component reward.

04

Testing shortcuts in the reward

Added adversarial tests targeting reward shortcuts and insufficient evidence gathering.

05

Advancing an adaptive adversary

Built WarRoom with an adversary that advances during an investigation and adapts to containment actions.

Read the system flow

Agent evaluation environments · Public explanation of the implementation boundaries.

  1. ObserveInspect the system and the current case state.
  2. Gather evidenceUse investigation tools and cross-reference the findings.
  3. ActClassify and remediate, or contain the simulated incident.
  4. EvaluateScore the actions and evidence, not only the final answer.

Compliance Auditor

A procedural compliance investigation with MCP tools and a multi-component reward. Tests challenge skipped investigation, invented findings, incorrect classification, and missing remediation.

Environment and reward logic ↗

Adversarial checks

Reward tests ask whether an agent can score well by guessing, flooding the evaluator with findings, or skipping evidence. Procedural cases vary scenarios so repeating one memorized action sequence is insufficient.

Reward shortcut tests ↗

WarRoom

A simulated security operations environment. The adversary advances during the investigation and responds to containment, so the sequence of actions changes the case.

WarRoom implementation ↗

Validation scope

Public projects built for the Meta PyTorch OpenEnv Hackathon 2026. No placement, benchmark score, or real-world security and compliance performance is claimed.

Hackathon projects

Inspect
the work.

Public simulation environments built for the Meta PyTorch OpenEnv Hackathon 2026. The repositories contain environment state, investigation tools, reward logic, and adversarial tests.

Next project: Repository intelligence ↗

Source material

Compliance environmentWarRoom environment
Public hackathon implementations. Simulated outcomes do not establish real-world security or compliance performance.

Opening the project index…

Browse all work ↗

Interactive exhibit