simulation
Your agent’s digital twin meets synthetic personas, reads synthetic documents and calls mocked services, covering every agentic path to find the gaps in correctness and performance, and the fixes that close them. Interactively or in CI.
/plugin marketplace add plurai-ai/plurai-plugins/plugin install plurai-simulation@plurai-plugins/reload-plugins/plurai-simulation:simulatecopilot plugin marketplace add plurai-ai/plurai-pluginscopilot plugin install plurai-simulation@plurai-plugins# Restart GitHub Copilot CLI so it loads the Plugin
/plurai-simulation:simulatecodex plugin marketplace add plurai-ai/plurai-pluginscodex plugin add plurai-simulation@plurai-plugins# Restart OpenAI Codex so it loads the Plugin
$simulate
Every fixed failure joins the regression loop. Production traces keep the simulation hyper-real, so issues from production stay blocked and new ones from development never ship.
Plurai root-causes each failure, separates real agent bugs from noise, and lands and verifies the fix, with the evidence attached, ranked and pointing at the exact code location.
Your agent can take the right action and still fail. Reading the output would never have told you.
Each run happens in an isolated sandbox. Nothing touches production, and the world resets to the same state every time, so you can measure how consistently your agent behaves across runs. Every score comes with its reasoning, in your repo and in the observability tools you already use.
39passed
3failed
Plurai models the full space of what your agent can do as a probabilistic flow graph, then generates the scenarios to cover it: hard personas that push back and probe your policies, and the tasks most likely to break it. Not a hand-picked set that hopes it caught the important cases.
Plurai reads your agent as built, maps your knowledge bases, decides what stays live and what gets mocked, and drops it into a simulated world of services, documents and world state. No integration work on your side.
RAG · retrievalFare rules & Baggage policy KB
TravelAssistantAgentLangGraph · TravelAssistantStateSkill1
Itinerary-planningTools4+
search_flightsget_bookingget_flight_statuscheck_baggageSub-agents2
FlightBookingAgentRefundAgent
Harnesssimulate.py
Claude code
Travel-Assistant
Ready
Plurai reads your agent as built, maps your knowledge bases, decides what stays live and what gets mocked, and drops it into a simulated world of services, documents and world state. No integration work on your side.
Plurai models the full space of what your agent can do as a probabilistic flow graph, then generates the scenarios to cover it: hard personas that push back and probe your policies, and the tasks most likely to break it. Not a hand-picked set that hopes it caught the important cases.
Each run happens in an isolated sandbox. Nothing touches production, and the world resets to the same state every time, so you can measure how consistently your agent behaves across runs. Every score comes with its reasoning, in your repo and in the observability tools you already use.
Plurai root-causes each failure, separates real agent bugs from noise, and lands and verifies the fix, with the evidence attached, ranked and pointing at the exact code location.
problem
Your agent has endless ways to fail
Every persona, workflow step, tool, document and world state multiplies the paths. Hand-written tests barely scratch what production throws at it.
Building a consistent, managed testing world is hard
Building the agent was easy. Proving it works outside the lab means keeping its world consistent across services, databases and tools. Hand-written evals are a drop in the ocean and barely move your coverage.
Even when you find an issue, fixing it is a gamble
Patch one failure and two more surface where you weren’t looking. Nothing tells you the fix actually held — or just pushed the problem somewhere you’re not watching.
Fixing production issues takes luck
Replaying one means rebuilding the exact world state it happened in. And agents are inconsistent, so even a perfect replay may never hit the failure again.
solution
RAG · retrievalFare rules & Baggage policy KBTravelAssistantAgentLangGraphTools4+
Topology
approved before the mock environment is deployed
Plurai wires your agent’s digital twin into a fully mocked world: Gmail, Slack, Stripe, Zendesk, your APIs and databases, plus every input it reads, from PDFs to email threads to tickets. Integrate production traces to turn that realistic world hyper-real. It all runs in an isolated sandbox, and the plugin does all the wiring, with no work on your side.
34% · 1.6 bits0.9% · 6.8 bits · hardestDimension graph19 states47 weighted transitions
You don’t need 300 tests. They’re expensive, and most don’t move the needle. Plurai spreads scenarios across your agent’s capabilities by weight: the common paths that carry your traffic, and the edge cases where serious failures and complex tasks hide.
What your users actually doshare of sessions
Trace briefoptionalcarried into generation
With traces or without them, Plurai finds what matters to test. It draws on whatever you have, from production traces and knowledge bases to PRDs and policies, and builds high-accuracy, cost-effective evals around it. Already in production? Pull as few as 20 traces through Plurai’s connector and the world turns hyper‑real.
refund_after_missed_connection
prompts.pybooking_api.pyFix planranked, with exact code locations
Every critical gap comes with its root cause, the evidence, and the exact file and lines to change. Production issues replay from the exact turn they broke, and every fix is validated across the full regression set, so fixing one thing doesn’t break two others. Steer it from your coding agent, let it run on its own, or run it headless in CI on every PR.
Start in your coding agent
Plurai does the heavy lifting on both sides: it wires the digital twin and brings the intelligence that decides what to test and how to judge it. With post-training, specialized models take your agent’s capabilities further at a fraction of the inference cost. It gets better every run while others are still writing test sets by hand.
task_outcome_quality · poor · 0.00
fixed · re-proven