TryAgent

Your agents act. You stay in control.

Guard high-stakes tool calls with one SDK wrapper. TryAgent pauses before the action, routes the decision with full context, and continues the moment your reviewer approves.

Start freeDocumentation

100 escalations free each month. No credit card.

agent runtimehuman gate active

agent calls payments.refund

guarded· escalation esc_4821 created

routed· Payments Ops · SLA 30m

approved· j.lee

executed· refund complete · run resumed

Drops into the stack you already run

LangGraphOpenAIAnthropicVercel AI SDKSlackWebhooks
Scoped API keysSigned callbacksTenant isolationSLA-backed routing

One SDK wrapper

guard() puts a human checkpoint around any tool or function without changing its core logic.

Pauses, then resumes

The tool waits and the exact workflow continues the moment your reviewer decides.

Production controls included

Routing, SLAs, decision history, and security ship with the primitive.

How a decision flows through TryAgent.

TryAgent sits between your agent and the people who own the risky calls. It routes each decision to the right reviewer, resolves it under an SLA, and resumes the exact paused run — so nothing hangs and nothing slips through.

  1. 01

    Escalate

    A guarded call sends the question, evidence, choices, run id, and policy key in one request.

  2. 02

    Route

    Policy, schedule, destinations, SLA, and timeout decide who receives it and what happens next.

  3. 03

    Resolve

    A person approves, edits, or rejects with the proposed action and evidence in view.

  4. 04

    Resume

    The guarded tool executes after approval, or a signed callback continues the exact paused workflow.

One guarded call

The SDK is the front door.

Wrap a tool or function with guard(). Its policy registers on the first escalation, the call waits for a human decision, and the original function only runs after approval.

guard.ts
const tryagent = new TryAgent({ apiKey: process.env.TRYAGENT_API_KEY }); const guardedRefund = tryagent.guard(  "payments.refund",  refund,); const result = await guardedRefund({ amount: 480 });// executes only after approval
Escalation lifecycle
  1. createdrun pauses
  2. openfinance on-call
  3. decidedchoice recorded
  4. resume callback deliveredrun continues

fail closed

No approval, no side effect.

A rejection or SLA timeout returns a structured denial your agent can handle. The protected tool never executes by accident.

The reviewer's view

Every reviewer decides with the full picture.

An escalation is never a vague ping. The reviewer sees the proposed action, the evidence behind it, the exact choices, and a recommendation drawn from how your team decided similar cases — so the call takes seconds, not a research project.

  • The proposed action, evidence, and exact choices — structured, not freeform.
  • A recommendation learned from your resolved escalations, with confidence in plain view.
  • Prior outcomes ride along, so reviewers know when to trust, edit, or reject.
TryAgent · InboxOpen 3Resolved
  • billing-copilotopen

    Refund $480 to customer #4821?

    recommended · approve 86%

  • support-agentopen

    Cancel the enterprise plan for Acme Inc?

    finance on-call · SLA 2h

  • ops-agentopen

    Delete 12k stale records from prod?

    compliance · awaiting review

  • billing-copilotdecided

    Refund #4771 to customer #3902

    approved · j.lee

billing-copilot · run_7f3open

Refund $480 to customer #4821?

evidence

Customer charged twice on Jun 14
Within the 30-day refund window
No prior refunds on this account
86%

Recommended

Approve

4 of 5 similar refunds approved by finance

ApproveEditReject
signed callback resumes the run

Control plane

The mechanics around the decision are already built.

It looks like a queue and a webhook until you need routing, permissions, schedules, signed callbacks, audit, timeouts, and learning to line up under production load.

Policy routing

Versioned policies route by agent, subject, amount, metadata, or any field your workflow sends.

Schedules and fallbacks

On-call ownership, escalation chains, SLAs, and timeout defaults keep runs from hanging.

Delivery where teams work

Send decisions to the TryAgent inbox, Slack, email, webhooks, or your own review surface.

Decision history

Every route, timeout, decision, delivery attempt, and resume callback is recorded.

Production security

Scoped keys, signed callbacks, tenant isolation, and reviewer permissions ship with the primitive.

Reviewer context

Show the proposed action, structured evidence, choices, recommendation, and prior outcomes together.

The payoff compounds

Every decision makes the next one faster.

A plain review queue costs the same on day 100 as day one — every decision starts from zero. TryAgent does not. Resolved outcomes become precedent, recommendations sharpen, and the next reviewer starts with the context of how your team handled similar calls.

  • Precedent is scoped to your workspace, agent, and policy — never shared across tenants.
  • Confidence and prior outcomes ride along on every escalation, so reviewers decide in seconds.
  • Recommendations stay advisory. Every guarded call still waits for a person to decide.
Recommendation confidence by week
similar outcomes foundreviewer still decides
Week 1
8%
Week 2
22%
Week 3
39%
Week 4
54%
Week 6
66%
Week 8
73%

Illustrative. Precedent improves recommendation confidence while a person remains responsible for every guarded call.

Put a human on your first high-stakes call today.

Start free with 100 escalations a month. Keep every risky agent decision resumable, auditable, and useful to every run after it.