Stress-test your agent before deploy.
Adversarial counterparties, prompt injection fuzzing, goal-drift scenarios.
iAgentSim will be the sandbox environment for testing agent behavior under adversarial conditions before deployment. Synthetic counterparty agents that attempt prompt injection, byzantine actors, resource pressure, and goal-drift scenarios. Cluster 3 roadmap 2028.
iAgentSim will provide four primitives
Adversarial counterparties
Pre-built red-team agents that attempt prompt injection, output manipulation, and context poisoning. Fuzz your agent before real attackers do.
Scenario library
Budget pressure (force your agent to overspend), byzantine oracle (poisoned input data), goal drift (gradual prompt shift), social engineering simulations.
Regression tracking
Run the same scenario suite after every prompt change or model upgrade. Catch behavioral regressions.
Replay + fork
Replay real iAgentLog traces. Fork at any decision point and ask 'what if the counterparty had been malicious?'
The gap iAgentSim fills
You cannot catch adversarial behavior in unit tests
Traditional testing gives known inputs and checks outputs. Adversarial testing injects behaviors that real-world attackers will try, which unit tests never imagine.
LLM eval ≠ agent eval
LangSmith datasets, Braintrust evals are great for LLM output quality. Bad for adversarial agent-to-agent interaction testing.
Integrates with iAgentLog and iAgentSafe
Pull real behavior baselines from iAgentLog, validate iAgentSafe policies actually fire in adversarial scenarios.
Planned endpoints in Roadmap 2028 (Cluster 3)
Preview of the planned API surface. OpenAPI 3.1 specification at /.well-known/openapi.yaml. Endpoints at api.iagentsim.com will serve requests at roadmap 2028 (cluster 3); agent-consumable JSON by design.
POST /v1/simulate |
Run adversarial scenario against your agent |
GET /v1/scenarios |
Catalog of available adversarial patterns |
POST /v1/replay |
Replay historical trace with adversarial injection |
GET /v1/results/{run_id} |
Retrieve simulation results |
iAgentSim is one layer
Sixteen products. One stack. One entity. Trust, discovery, observability, payments, safety, simulation, composition, memory, identity, legal, markets, and owned compute underneath. Each layer reinforces the others. Use one or use them all.