AI workflows are glue.
Glue breaks.

AGNT5 adds the durability, retries, and state needed to handle queues, callbacks, and human handoffs — so your workflows and agents recover instead of fail.

Start building View docs

triage_ticket.py

# Durable workflow with human approval

from agnt5 import durable, Context

@durable('triage_ticket', retries=5)

async def triage(ctx: Context, ticket: dict):

# Step 1: Draft a response using LLM

payload = {'role': 'user', 'content': ticket['text']}

draft = await ctx.step('draft', lambda: ctx.llm.chat(messages=[payload]))

# Step 2: Wait for human approval (can pause for days)

decision = await ctx.signal.wait('agent_approval', timeout='24h')

if decision != 'approve':

return 'held'

# Step 3: Send the email

await ctx.step('send', lambda: ctx.email.send(

to=ticket['user'],

body=draft.text,

))

return 'sent'

Waiting to start...

[01] THE PROBLEM

The "glue" is harder than you think.

Solving distributed systems challenges in application code is a liability. AGNT5 handles the complexity of the network and infrastructure for you.

A deployment happens while an agent is waiting for an LLM response. In standard code, that process dies and the context is lost forever.

run_agent.py

# WITHOUT AGNT5 (The "Manual State" Nightmare)
async def run_agent(run_id):
    state = db.get_run(run_id)

    if state.step == 'start':
        try:
            # If server crashes here, result is lost
            result = await llm.generate()
            # If DB commit fails here, you might re-run LLM later
            db.update(run_id, step='processed', data=result)
        except SystemExit:
            # Hope your graceful shutdown works...
            db.save_checkpoint(run_id)

> Waiting to start...

[02] PLATFORM PRIMITIVES

Production primitives for AI workflows.

Everything you need to build reliable AI applications, built into the runtime.

Durable Runtime

Survive crashes, deploys, restarts. Your workflow picks up exactly where it left off. No manual checkpointing required.

Retries & Schedules

Declarative retry policies, not scattered try/catch. Cron schedules built-in. Set once, run forever.

Human Handoffs

Wait for approvals as first-class steps. Pause for days, resume instantly. No polling, no webhooks.

Observability & Evals

Traces, logs, metrics, eval hooks. Debug in development, monitor in production. OpenTelemetry built-in.

[03] DEVELOPER WORKFLOW

How AGNT5 fits into your stack.

From local development to production deployment in three steps.

Step 01

Install

terminal

$ pip install agnt5

Install the SDK. Define workflows in your existing Python codebase. No infrastructure to set up.

Step 02

Develop

terminal

$ agnt5 dev

Run locally with full visibility. Test workflows, inspect state, debug with traces.

Step 03

Deploy

terminal

$ agnt5 deploy

Deploy to your infra or AGNT5 Cloud. Same code, same guarantees. Scale automatically.

Install--->Develop--->Deploy

[04] USE CASES

Built for AI workflows that need to work.

Real applications running in production today.

Customer Support Agents

Triage tickets, draft responses, wait for approval, send. All durable. No lost context between handoffs. Handle escalations that span days without losing state.

human-in-loopmulti-stepasync-approvals

Multi-step RAG Pipelines

Chunk, embed, store, retrieve, generate. Decoupled compute with recoverable steps.

Data Processing Pipelines

ETL, transformations, exactly-once. Never double-process a record.

Approval Workflows

Route to reviewers, collect signatures, handle escalations. Pause for days or seconds, resume instantly. No polling, no webhooks, no manual state management.

Start building reliable AI workflows today.

Get started in minutes. No infrastructure to manage.

Get started View on GitHub

AI workflows are mostly glue . AGNT5 makes the glue reliable.

Integrations, queues, callbacks, human handoffs, error paths — AGNT5 adds durability, state, retries, traces, and evals so your workflow survives real production.

Start Project View Docs

workflow.py

from agnt5 import workflow, task

@task(retries=3, backoff="exponential", timeout="30s")
async def call_llm(prompt: str):
    # If OpenAI 500s or rate limits, we automatically retry
    return await openai.chat.completions.create(...)

@workflow
async def process_document(doc_id: str):
    # State is persisted automatically after every step
    text = await extract_text(doc_id)
    summary = await call_llm(f"Summarize: {text}")

    # If this fails, we resume right here. No re-running the LLM.
    await save_to_db(summary)

Without AGNT5

× Manual try/catch loops in every function.
× Writing custom backoff logic (sleep/retry).
× If the script crashes, you restart from scratch.
× Paying for the same LLM tokens twice on restart.

With AGNT5

✓ Declarative retry policies (1 line of code).
✓ Durable Execution: Resumes exactly where it failed.
✓ Zero wasted tokens on replays.
✓ Timeout handling out of the box.

workflow.py

from agnt5 import workflow, activities

@workflow
async def content_generator(topic: str):
    # Step 1: Generate Draft
    draft = await activities.generate_blog_post(topic)

    # Step 2: Pause for Human Review
    # The workflow suspends here. It consumes 0 compute while waiting.
    # It can wait for minutes, days, or months.
    approval = await workflow.wait_for_signal("user_approval")

    if approval.granted:
        await activities.publish_post(draft)
    else:
        await activities.discard_post(draft)

Without AGNT5

× Building a custom database state machine (status="pending").
× Setting up cron jobs to poll for updates.
× Writing webhook handlers to wake up the process.
× Managing complex race conditions.

With AGNT5

✓ Pause & Resume logic is just one function call.
✓ No polling or cron jobs required.
✓ State is automatically hydrated when the human clicks "Approve."
✓ Workflows can sleep for months practically for free.

workflow.py

from agnt5 import workflow, check
from pydantic import BaseModel

class UserInfo(BaseModel):
    name: str
    age: int

@workflow
async def extract_data(email_text: str):
    # Step 1: Structured Extraction
    # Automatically validates JSON against Pydantic schema
    user_data = await agnt5.chat(
        model="gpt-4",
        response_model=UserInfo,
        max_retries=2
    )

    # Step 2: Semantic Evaluation
    # Fails the workflow if the confidence score is too low
    await check.assert_true(
        agnt5.evals.hallucination_score(user_data) < 0.1,
        on_fail="raise_error"
    )

    return user_data

Without AGNT5

× Writing Regex to parse broken JSON output.
× Manually implementing "correction prompts" when parsing fails.
× No visibility into hallucination rates in production.
× Bad data silently polluting your database.

With AGNT5

✓ Guaranteed JSON output matching your schema.
✓ Auto-healing: AGNT5 asks the LLM to fix its own errors.
✓ Built-in Evals run on every request.
✓ Production-grade safety nets.

AI Workflows from Dev to Prod, Simplified

AGNT5 enables development teams to build production-grade AI workflows without the operational complexity. Write workflows as code, deploy to any environment, and monitor everything with full observability—from laptop to production.

AGNT5 Orchestration holds it all together

Build workflows that survive failures, scale across teams, and maintain full observability. AGNT5's orchestration layer handles state, retries, and coordination—so you can focus on building features.

Durable execution with automatic state management
Built-in retries and idempotency guarantees
Replay and debug historical executions

Screenshot / Diagram Placeholder

AGNT5 Handles It

Stop building infrastructure. Focus on your AI workflows while AGNT5 handles the operational complexity.

State Management

Automatic checkpointing and state persistence across restarts
Retry Logic

Built-in exponential backoff and configurable retry policies
Observability

End-to-end traces with OpenTelemetry integration
Human-in-the-Loop

Built-in signals and approvals for agent workflows
Deploy Anywhere

Run locally, self-host, or use AGNT5 Cloud—same code everywhere

Developers trust AGNT5 for production AI workflows

Built for teams who ship AI agents to production. AGNT5 provides the reliability, observability, and developer experience you need to move fast with confidence.

ANTHROPIC

Coinbase

Microsoft

Perplexity

The Easiest Way to Build AI Agents

AGNT5 removes the complexity from production AI workflows. Focus on building features while we handle the infrastructure.

Start Fast

Install the CLI, write your first workflow, and run it locally in under 5 minutes. No infrastructure setup required.

Get started

Deploy Anywhere

Run locally, self-host, or use AGNT5 Cloud. Same code works everywhere with zero configuration changes.

Learn more

Monitor Everything

Full observability with OpenTelemetry integration. Track every step, replay failures, and debug with confidence.

Learn more

Scale Confidently

Built-in state management, automatic retries, and durable execution. Your workflows survive failures and scale horizontally.

Learn more

Build in Your Language

Native SDKs for Python and TypeScript. Write workflows in the language you already know and love.

View SDKs

Focus on Logic

Stop building infrastructure. AGNT5 handles queuing, state, retries, and coordination so you can focus on business logic.

Learn more

AI workflows are glue. Glue breaks.

The "glue" is harder than you think.

Production primitives for AI workflows.

Durable Runtime

Retries & Schedules

Human Handoffs

Observability & Evals

How AGNT5 fits into your stack.

Install

Develop

Deploy

Built for AI workflows that need to work.

Customer Support Agents

Multi-step RAG Pipelines

Data Processing Pipelines

Approval Workflows

Start building reliable AI workflows today.

AI workflows are mostly glue . AGNT5 makes the glue reliable.

Without AGNT5

With AGNT5

Without AGNT5

With AGNT5

Without AGNT5

With AGNT5

AI Workflows from Dev to Prod, Simplified

AGNT5 Orchestration holds it all together

AGNT5 Handles It

State Management

Retry Logic

Observability

Human-in-the-Loop

Deploy Anywhere

Developers trust AGNT5 for production AI workflows

The Easiest Way to Build AI Agents

Start Fast

Deploy Anywhere

Monitor Everything

Scale Confidently

Build in Your Language

Focus on Logic

AI workflows are glue.
Glue breaks.