Skip to content

Make your long-running agents reliable.

AGNT5 runs your agent and records every step. Crashed runs resume where they stopped, evals tell you if a deploy is better or worse, and bad runs become tests you replay against your next change.

Start for free. No credit card required.

Your agent, defined in code.

Write agents, tools, and workflows in Python, TypeScript, or Go. The highlighted calls are AGNT5; everything else is your own code.

Read the build docs
src/support/workflows.py
from agnt5 import WorkflowContext, workflow
from support.agents import support_agent
from support.functions import issue_refund


@workflow
async def handle_refund(ctx: WorkflowContext, order_id: str, message: str) -> dict:
    # every model and tool call is recorded in the run's trace
    answer = await support_agent.run(f"Order {order_id}: {message}", context=ctx)

    # pauses here for hours or days, until someone answers
    decision = await ctx.wait_for_user(
        question=f"Refund order {order_id}?\n\n{answer.output}",
        input_type="approval",
        options=[{"id": "approve", "label": "Approve"},
                 {"id": "reject", "label": "Reject"}],
    )
    if decision == "reject":
        return {"status": "declined"}

    # checkpointed: once it succeeds, it never runs again
    await ctx.step(issue_refund, order_id)
    return {"status": "refunded"}

Next: deploy support-desk and watch handle_refund run

Run it, and watch every step.

Deploy support-desk with agnt5 deploy, and every call to handle_refund becomes a run you can open: each model call, tool call, retry, and approval, as it happens.

Read the run docs
handle_refund · run 7f3aCompleted1h 12m
Input
{"order_id": "A-1042", "message": "My blender arrived cracked."}
Output
{"status": "refunded"}
Deployment
deploy 14 · production
Started by
agnt5 run handle_refund --env production

Code workflows.py:7workflows.ts:6workflows.go:9

Next: fix the runs that broke the policy

Fix it on the runs that failed.

Scorers grade live runs of handle_refund. When some fail, AGNT5 groups them into one issue, points at the code, and lets you test your fix on those same runs before it ships.

Read the improve docs
Online eval · follows_refund_policybelow 95% floor

91%passed this week

256runs scored (10% sample)

23failed

97Mon93Tue90Wed92Thu89Fri91Sat88Sun95% floor
  • run 51d0Recommended a $129.00 refund on final-sale order A-3381
  • run 7f3aRefund matches §2, damaged on arrival
  • run a9c4Recommended a $74.00 refund on final-sale order A-3290
  • run e04bRefund matches §3, unopened
  • run 77e2Recommended a $210.00 refund on final-sale order A-3402

Deploy 15 goes back to Run, where its runs are traced and scored

Bring the agents you already have.

No rewrite. Wrap your agent’s entry point in an AGNT5 function and call your framework exactly as you do today. Your prompts, tools, and model providers stay as they are.

  1. Step 1

    Install the CLI

    curl -LsSf https://agnt5.com/cli.sh | bash

    For Python, TypeScript, or Go projects.

  2. Step 2

    Wrap your existing agent

    @function # on your entry point

    One decorator. The agent code inside it doesn’t change.

  3. Step 3

    Deploy

    agnt5 deploy

    Your first run is recorded, start to finish.

Works with

OpenAI Agents SDKClaude Agent SDKGoogle ADKVercel AI SDKCloudflare Workers

Deployment

Run it in our cloud or yours.

Use AGNT5 fully managed, or run it inside your own environment. If prompts, outputs, and customer data can’t leave your network, they don’t have to.

AGNT5 control plane
Command centerDeployments
Your network
Your agentsRuntimeRun recordsPromptsOutputsCustomer data
  • Fully managed

    Sign up and deploy. We run everything.

  • In your cloud

    Your agents run, and your data stays, in your environment.

  • Same features either way

    Build, Run, and Improve work the same in both, deploy verdicts included.

Deploy your first agent in 5 minutes.

Start free with 250K steps a month.

Enterprise

Request a demo

Tell us what you're building or how we can help. We'll reply by email.