TryAgent

About TryAgent

TryAgent puts a human checkpoint in front of the tool calls an AI agent should not make unsupervised. You wrap a function with guard() from the SDK and nothing else about the tool changes; the first time it is invoked, its policy registers, the call pauses, and TryAgent routes the decision to whoever owns that risk with the proposed action, the supporting evidence, the exact choices available and the run id attached. A reviewer approves, edits or rejects, and the original function executes — or, if the workflow was resumed asynchronously, a signed callback continues the exact paused run. It fails closed: a rejection or an SLA timeout returns a structured denial the agent can handle rather than a hang, and the protected tool never fires by accident. Around that primitive sits the control plane you would otherwise build yourself — versioned policies that route by agent, subject, amount or any metadata field, on-call schedules and escalation chains, scoped API keys, signed webhooks, tenant isolation, a tamper-evident audit chain, decision exports and SIEM streaming. The interesting design choice is that resolved escalations become precedent: the reviewer's view carries a recommendation with a confidence figure drawn from how the team decided similar cases, and rubrics can graduate a decision boundary to autonomy with automatic demotion. It drops into LangGraph, the OpenAI and Anthropic SDKs, the Vercel AI SDK, Slack and plain webhooks.

Key Features

Tags