Kashikoi is a no-code simulation platform for evaluating AI agents, built by the ex-Moveworks lead behind 250+ shipped enterprise agents.
Kashikoi is built by a 2-person founding team for ML engineers and AI product leads who need to catch agent bugs before customers do. It replaces ad hoc prompt tuning and generic public leaderboards with simulations tailored to a team's own AI stack, though pricing requires booking a demo.
Builds a full benchmark from a single prompt with no code, compressing what competing eval platforms need engineers to configure by hand.
The founding CTO's Moveworks Simulation and Evaluation background is a real edge, but multi-turn agent simulation is a crowded 2026 category (Archal, Openlayer, Galileo, Maxim AI) with no disclosed technical moat like a patent or proprietary dataset.
Main trade-off
No public pricing, no disclosed named customers, and a 2-person team as of 2026, all standard early-YC risk signals for a buyer evaluating vendor stability.
No public pricing as of 2026. Kashikoi runs on a request-a-demo model: book time via Calendly on getkashikoi.com or email founders@getkashikoi.com to get a quote.
Launched from Y Combinator's Spring 2025 batch, Kashikoi is a simulation engine that runs autonomous, multi-turn conversations to benchmark AI agents before they reach production, replacing static public benchmarks and manual prompt tuning with structured behavioral assessments from custom world models.
Autonomous World-Model Simulation
Generates CPU-friendly world models that run multi-turn conversations with the agent under test on their own, producing a behavioral assessment without a human writing evaluation prompts.
Single-Prompt, No-Code Setup
Builds a full benchmark from one prompt, replacing the manual work of authoring eval scripts or rubrics for each agent.
Custom AI-Stack Integrations
Connects into a team's existing agent stack through purpose-built integrations rather than a fixed connector list.
Automatic Prompt Optimization
Tunes the agent's prompts based on gaps the simulation surfaces, cutting down the manual prompt-tuning loop.
Regression and Staleness Detection
Flags an eval suite once it stops reflecting the agent's current behavior, so a swapped model or prompt doesn't silently pass a stale test.
Most users expect these and they are not currently supported.
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| Quote | No | Web | -- | |
| -- | Yes | Web | 3.4 | |
| -- | Yes | Web | 3.9 | |
| -- | Yes | Web | 3.0 |
Proprietary
There is no published price list. A prospective customer books a demo via Calendly on getkashikoi.com, or emails founders@getkashikoi.com, and the team quotes a custom price from there.
There is no self-serve free tier. The only no-cost step is booking a demo call or emailing the founders directly to discuss access, so there is nothing to try before that conversation happens.
Archal runs evals against sandboxed clones of tools like GitHub and Stripe. Openlayer pairs a large pre-built test suite with EU AI Act compliance mapping. Both are more built-out than Kashikoi's single-prompt, custom-integration approach as of 2026.
Openlayer runs 175+ pre-built evaluation tests with real-time guardrails and EU AI Act mapping out of the box. Kashikoi instead builds a bespoke benchmark from a single prompt and custom AI-stack integrations, trading breadth for a setup with no coding required.
Book a demo through Calendly on getkashikoi.com or email founders@getkashikoi.com. The team then connects a custom integration into your AI stack and builds an initial benchmark from a single prompt, with no code required on your side.
Kashikoi is a simulation engine for benchmarking AI agents, founded in 2025 by Aaksha Meghawat and Tim Michaud and backed by Y Combinator's Spring 2025 batch. The company was built to replace two common but weak ways of testing an AI agent: manual prompt tuning and scoring against static public benchmarks that rarely match a real production conversation.
The core mechanism is a set of CPU-friendly world models that autonomously interview the agent under test across multi-turn conversations, then produce a structured behavioral assessment instead of a single pass or fail score. Setup starts from one prompt with no code required, and the platform builds custom integrations into a team's existing AI stack so the simulation runs against the real agent rather than a mock. Meghawat led the Simulation and Evaluation stack at Moveworks (ServiceNow), where the team shipped 250+ customized enterprise agents to Fortune 500 and federal customers, and Kashikoi extends that internal approach into a standalone product.
The platform is aimed at ML engineers and AI product teams shipping support bots, data agents, and coding assistants who need to catch behavioral regressions before a release, plus teams that want benchmarks aligned to their own product's success criteria rather than a generic leaderboard. It also automatically tunes prompts based on the gaps a simulation surfaces, and flags when an existing benchmark has gone stale as the underlying agent or model changes.
Kashikoi has no public pricing page as of 2026. Access starts with a demo booked through Calendly or a direct email to founders@getkashikoi.com; the product itself is browser-based only, with no desktop or mobile client. As a 2-person team with no disclosed named customers or compliance attestations yet, it suits teams comfortable evaluating an early-stage vendor directly rather than through a self-serve trial.
Product facts on this page (pricing, plans, integrations, compliance) are read from Kashikoi's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.