Skip to content
Tether

For AI agents, models and robots

Crash tests for AI agents.

See what your agent will do before it does it. Block it if it's dangerous. Undo it if it slips.

Free to start. Bring any model or agent. Run your first crash test in a few minutes.

A test robot rides a sled toward a wall. The Tether gate drops in front of it and stops it, and a mallet bonks it on the head.
sandbox / crash testillustration

task>

    Undorefund reversed with payments.recharge

    The problem

    AI agents now send money, email customers and move robots. Nobody crash-tests them first.

    Cars are crash-tested before they reach the road. Agents go straight to production, and the first sign of trouble is a real customer.

    Example: one timeout, two refunds
    1. 09:00:01Agent calls payments.refund for order 4821, $120.
    2. 09:00:31The payment provider times out. The refund actually went through.
    3. 09:00:32The agent thinks it failed, and tries again.
    4. 09:00:33Without a gate: the customer is paid twice. Nobody notices for a week.
    5. 09:00:33With Tether: the second refund is blocked as a duplicate, with the reason on record.

    How it works

    Contract. Gate. Replay.

    Three small ideas that together make an agent's actions predictable, stoppable and reversible.

    Contract01
    payments.refund@1
    reversibility: compensable
    max amount: 1000
    needs a human: amount > 200
    no repeats within: 24h

    Every action says what it does.

    Each tool your agent can use gets a short contract: its effects, how far it can reach, and whether it can be undone, made right, or never taken back.

    Gate02
    intentgateallowblockhold

    Every action passes one gate.

    The gate checks the contract, live state and your rules, then allows, blocks or holds for a human. It decides from facts, never from what the agent says.

    Replay03

    Every new version is crash-tested first.

    Before a new prompt or model ships, it reruns twenty trap scenarios in a sandbox copy of your business, several times over, and gets a risk score.

    What you can crash-test

    Models, agents and robots. One lab.

    Each crash test runs inside a sandbox: a fake shop with customers, orders and an inbox, plus a simulated robot arm. Nothing real is ever touched.

    Models

    Point Tether at any Claude model or any OpenAI-compatible endpoint. It plays a shop operations agent and walks into every trap.

    claude-sonnet-5-5, your fine-tune, a local model behind an API

    Agents

    Paste your agent's system prompt, pick its model, and test each new version against the last one before it ships.

    support agents, refund bots, sales and ops assistants

    Robots

    A simulated arm goes through the same gate as a refund. Safe zones, speed near people, unknown objects and sensor loss.

    pick and place, workcells, anything with a safe zone

    The scoreboard

    Public crash-test ratings for AI agents.

    Every score comes from real runs with a published method, repeated several times, with the spread shown. Never rounded up, never cherry-picked.

    All ratings

    demo data These rows are made-up examples that show the format. Real ratings replace them as runs are published.

    Agent or modelRatingScoreShare
    Demo agent A (refund desk)
    demo datatraps-v1 x5
    pass
    91.5
    88 to 95
    Demo agent B (support inbox)
    demo datatraps-v1 x5
    warn
    72.0
    66 to 79
    Demo agent C (warehouse arm)
    demo datatraps-v1 x5
    fail
    48.0
    40 to 55

    Software and physical

    A refund and a robot arm pass the same gate.

    A refund has a limit. An arm has a safe zone. To Tether both are contracts with a blast radius, checked before anything moves.

    API call

    GATE
    BLOCKED

    payments.refund(ord_4822, amount: 1500)
    blast_radius: amount 1500 is above the limit of 1000

    Robot arm

    GATE
    BLOCKED

    arm.move_to(x: 0.8, y: 0.1, z: 0.2)
    blast_radius: target is outside the safe zone

    Fits into what you already run

    A sandbox gate before production, in one HTTP call.

    Fintech, support desks, internal tools, robots: if your AI takes actions, put Tether in front of them. Any language, any framework, or MCP.

    1. 1Describe your actionsUse our contracts or write your own: a few lines per tool.
    2. 2Ask the gate firstOne HTTP call before each action. Allowed, blocked or held, with the reason.
    3. 3Crash-test every releaseRun the trap suite against each new prompt or model before it ships.
    Read the quickstart
    const decision = await fetch("https://tether.wamasoda.com/api/v1/intents", {
      method: "POST",
      headers: { Authorization: "Bearer tk_...", "Content-Type": "application/json" },
      body: JSON.stringify({
        contract: "payments.refund",
        input: { orderId: "ord_4821", amount: 120, currency: "USD",
                 destination: "original_method", reason: "Never arrived" },
        mode: "sandbox",
        probes: { "order.exists": true, "order.paid": true, "refund.within_remaining": true,
                  "refund.currency_matches": true, "refund.original_method": true },
      }),
    }).then((response) => response.json())
    
    if (decision.data.verdict === "allow") await refund()

    Who it is for

    Teams whose AI does things, not just says things.

    • Agent startups

      Show customers a crash-test report before they let your agent near their money or their inbox.

    • Fintech and payments

      Duplicate refunds, wrong currencies and new-card fraud are traps one through five.

    • Robotics teams

      Test the planner that drives your arm against safe zones and people, in simulation first.

    • AI researchers

      Compare models on the same traps with a published, versioned method and raw traces.

    The contract spec is open.

    Action Contracts are plain JSON Schema under the Apache 2.0 licence. Describe any action once, in any language, and every tool that reads the spec understands it. The hosted gate, ledger and Replay Lab are the product.

    Pricing

    Start free. Pay when it protects real money.

    Run 3 crash tests a day on us, or unlimited tests on your own model key. Early teams get hands-on help setting up.

    Free

    $0

    Everything you need to crash-test your first agent.

    • 3 crash tests a day on Tether's model key
    • Unlimited tests with your own model key
    • All 20 traps, risk score and full traces
    • Gate API, sandbox mode and shareable reports
    Start free

    Team

    Early access

    For teams putting an agent in front of real customers.

    • Live mode: the gate in front of production
    • Human approvals for held actions
    • Custom contracts and traps for your business
    • Version-to-version behaviour diffs
    Talk to us

    Enterprise and robotics

    Custom

    For regulated money flows and physical systems.

    • Private deployment
    • Robot adapters beyond the simulator
    • Risk reports for your board or regulator
    • Direct line to the founder
    Book a crash test

    Before your agent ships

    Find out where it breaks while it is still safe to break.

    Tell us what your agent does and we will crash-test it with you, including traps written for your business.

    Or just get updates on new traps and ratings