ancha.

Production, simulated.

Stateful testing playgrounds for AI agents.

We give an agent a working copy of the world it will act in: the tools, the data inside them and the people it talks to. Teams see how it behaves before it ships, and again on every change.

  • Slack
  • GitHub
  • Datadog
  • PagerDuty
  • Salesforce
  • HubSpot
  • Gong
  • GA4
  • Google Ads
  • LinkedIn Ads
  • ChatGPT Ads
Ancha · Pre-seedancha.io · October 2026

Agents now take real actions inside their customers' software.

Support agent SlackReads the question CustomerReplies mid-task HubSpotLooks up the account DatadogChecks the error rate GitHubFiles the bug PagerDutyPages on-call

Copying this world is a project of its own.

  • Vendor sandboxes are empty, capped or Enterprise‑only
  • Every tool, doc and vector DB needs realistic data, mess included
  • The tools must work together and reset after every run
  • Customers reply in ways nobody scripted

So teams end up testing on clean mock data or on their customers.

Introduction2

Agent engineering teams can't see how their agent will act in a real customer's environment until it's live.

Evals grade what the agent says. Mocks return clean, made-up data. Neither shows what it does with a real customer's tools, data and people.

Problem3

Coding agents learned to test web apps in a browser. Agents need a sandbox.

Web apps
  1. Coding agent builds the app
  2. Runs the tests
  3. Opens it in a browser
  4. Sees what broke and fixes it
  5. Ships in fewer iterations

Built into Cursor (Oct 2025), Claude Code (Feb 2026) and Codex (Apr 2026)

AI agents
  1. Coding agent builds the agent
  2. Runs tests and evals
  3. Runs it in an Ancha sandbox
  4. Sees what broke and fixes it
  5. Ships in fewer iterations

Isn't standardized today

Coding agents already run tests. Opening the app in a browser is what cut the iterations. A sandbox will do the same for agents.

Why now4

Ancha spins up a production‑like copy of your agent's world.

ancha sandbox Your agent REPLICASlackChannels, threads, events SIMULATEDCustomerReplies on a simulated clock REPLICAHubSpotContacts, companies, deals REPLICADatadogMonitors, logs, metrics REPLICAGitHubRepos, issues, PRs, checks REPLICAPagerDutyIncidents, on-call, paging Resets after every run
  • Tools

    A stateful twin of every tool, wired together over one state, with the same endpoints and credential shapes as the real ones.

  • Data

    Seeded records in every tool from versioned starter datasets. Production-scale synthetic data is next.

  • People

    Teammates and customers who post and reply on a simulated clock. Off-script replies are next.

  • CLI-first

    A coding agent creates, forks, inspects and resets worlds from the terminal. Nobody needs the UI.

Solution5

Live today: run an agent, read back every tool.

The Ancha console: attempt 1 of the demo on-call test, showing the PagerDuty incident the agent raised and acknowledged
  • app.ancha.io

    Hosted workspaces, one private machine per account. Sign in with email.

  • @ancha-labs/cli on npm

    Coding agents create, fork and inspect worlds from the terminal; the same key works in CI.

  • 11 tool twins

    Slack, GitHub, Datadog and PagerDuty verified; seven more in closed alpha.

  • Isolated playground runs

    Your agent in its own container, N attempts, every call and end state kept.

Attempt 1 of the demo on-call test: the incident the agent raised and acknowledged, read back from the PagerDuty twin. Next: checks and pass rates on the end state, then calibration on production traces.

Product6

Sandboxing becomes a core step in how every agent ships.

We believe it is a necessary part of building agents, the way nobody ships code without CI.

Build→Test in a sandbox→Ship→Run in production
  • In agent builders

    A sandbox step wherever agents get built

  • Before launch

    Prove a new agent works before customer one

  • On every change

    Catch regressions from prompts, models and tools

  • For risky actions

    Try refunds and deletes where nothing breaks

  • For each new customer

    Rehearse in a copy of their setup before go-live

  • For training, later

    The same worlds help models learn real work

Vision7

A $10B-a-year market by 2030.

  • TAM · $10B a year

    Testing every AI agent. What teams could spend on it each year by 2030.

  • SAM · $2.9B a year

    Business agents. Agents that work inside business software, built by enterprises and agent startups.

  • SOM · $90M a year

    Our share. What we can realistically win by 2030.

Not counted: AI labs, who buy the same kind of worlds to train models.

TAM: 20% of a $52.6B AI agent market in 2030 (MarketsandMarkets); testing took 26–35% of IT budgets in 2015–17 (World Quality Report). SAM: 20,000 enterprises at $125K and 10,000 agent startups at $40K a year. SOM: 400 of those enterprises and 1,000 of those startups.

Market8

Every other layer of agent testing is already funded and growing.

Competition9

We win on realism, proof and coverage.

  • RealismThe whole world, not one piece

    Agents act on tools, search docs and talk to people in the same task. We copy all of it, together.

  • ProofCalibrated on production

    Production traces tune the replicas and become new scenarios. What we sell is how well a sandbox pass predicts production.

  • CoverageEvery customer's setup, on every change

    One agent ships into many companies, each set up differently. Every change is tested against a copy shaped like each one.

This is the self-evolving part: the sandbox learns from production and gets closer to the real world with every run.

Competitive advantage10

Land through startups and builders, then grow with usage.

Direct

Startups building business agents

Partners

Inside the tools agents get built with, like Codex, Lovable and Emergent

Free tier

Vibe coders start free with limits, then pay as usage grows

Enterprise

Required before release

Later, not counted

AI labs, for training

A free tier with limits is how Vercel and Composio built awareness and usage. We follow the same playbook.

Go to market11

A live product, and early signals from agent teams.

Shipped

  • app.ancha.io · hosted workspaces, one private machine per account
  • @ancha-labs/cli on npm · sandboxes and secrets from the terminal
  • 11 tool twins · 4 verified, 7 in closed alpha
  • Isolated playground runs · two containers per attempt, every call and end state kept
  • Demo workspace · 5 agents and 6 tests across support, on-call, CRM and release notes

Signals

  • __agent teams interviewed
  • __design partners
  • __teams on the waitlist

“What an agent builder told us about testing their agent today”

Name, role, company
Traction12

Jayanth and Patrick are building Ancha.

  • Jayanth PuttaCo-founder & CEO
    • Built IBM's first LLM and agent observability product from the ground up, now in production at enterprises worldwide.
    • Built Kubernetes observability from 0 to 1 at OpsRamp, about $2M in new revenue; OpsRamp was acquired by HPE. Before that, product and solutions engineering at Datadog and firewall VPN engineering at Cisco.
    • Founded TourByte, an AI agent that answered renters and booked tours for NYC brokers. Open-sourced LoopEval, an eval framework for agents.
    • IBM
    • HPEOpsRamp
    • Datadog
    • Cisco
    • NIT Karnataka
  • Patrick LiuCo-founder & CTO
    • Founded SimpleFunctions: agents that researched events and traded live on Polymarket and Kalshi.
    • Built SparkCo Harness, an open-source CLI where a local and a Cloudflare agent runtime negotiate, share state and change their own environments. Shipped a phone agent that ran live at a skilled nursing facility.
    • MA in Physics, Brown. BA in Physics and Math at UW–Madison in two years. Hand-wrote a neural network to classify particles in CERN collider data.
    • SimpleFunctions
    • Brown University
    • University of Wisconsin–Madison
Team13

A pre-seed round to take Ancha from design partners to paying teams.

Raising $ __

Pre-seed, __ months of runway

Where it goes
  • __%
    Engineering

    Tool replicas, the synthetic data engine and evals

  • __%
    Go to market

    Design partners and builder integrations

  • __%
    Infrastructure

    Compute for sandboxes and test runs

Every agent should be tested in a copy of its world before it ships.

The askancha.io