# AGNT5

> One runtime for durable AI agent execution — checkpoint every step, trace every call, replay any run for evals.

Full documentation: https://agnt5.com/docs

## Core Concepts

- [Agents](https://agnt5.com/docs/concepts/agents.md): LLM-driven loops — instructions plus model plus tools, hosted inside step boundaries so their non-determinism is contained.
- [Architecture overview](https://agnt5.com/docs/concepts/architecture-overview.md): The map — Gateway, Engine, Coordinator, plus your workers — and how they fit together in a single binary.
- [Determinism — why workflows have rules](https://agnt5.com/docs/concepts/determinism.md): The contract on workflow code — replay must arrive at the same step calls in the same order — and how to keep your code on the right side of it.
- [Durable execution](https://agnt5.com/docs/concepts/durable-execution.md): The runtime guarantee that a workflow's progress survives crashes — completed steps are not re-run.
- [Event sourcing and replay](https://agnt5.com/docs/concepts/event-sourcing-and-replay.md): How AGNT5 records every step's input and output to a journal, and how replay reads the journal to skip work that already ran.
- [Functions](https://agnt5.com/docs/concepts/functions.md): Registered, addressable units of work — the smallest primitive in AGNT5 and the building block steps and tools both reach for.
- [Picking the right primitive](https://agnt5.com/docs/concepts/picking-the-right-primitive.md): Functions, workflows, agents, and tools — what each one is for, and which one to reach for when.
- [What the runtime owns vs. your code](https://agnt5.com/docs/concepts/runtime-vs-your-code.md): The responsibility boundary — what AGNT5 takes care of for you, and what stays in your code.
- [Sandbox isolation tiers](https://agnt5.com/docs/concepts/sandbox-isolation-tiers.md): How worker code is isolated — process, container, microVM — and the cost, latency, and security tradeoffs at each tier.
- [The improvement loop](https://agnt5.com/docs/concepts/the-improvement-loop.md): The trace → eval → edit → deploy cycle — the loop that turns durable execution into a product, not only a runtime feature.
- [Tools](https://agnt5.com/docs/concepts/tools.md): Capabilities agents can invoke during their loop — `@tool`-decorated callables, with idempotency and serializable returns as the disciplines that matter.
- [Versioning and deployment model](https://agnt5.com/docs/concepts/versioning-and-deployment.md): How AGNT5 handles workflow versioning — deployments are immutable, in-flight runs stay on their version, environments are pointers.
- [Workflows](https://agnt5.com/docs/concepts/workflows.md): Durable orchestrators — async functions whose progress survives crashes through journaled step boundaries.
- [Workflows, steps, and agents](https://agnt5.com/docs/concepts/workflows-steps-and-agents.md): The three primitives in AGNT5 — what they are, how they fit together, and which one you reach for when.

## Get Started

- [Build](https://agnt5.com/docs/get-started/build.md): Build your first workflow locally — workflows, steps, agents, and tools.
- [Build locally](https://agnt5.com/docs/get-started/build-locally.md): Use agnt5 dev to iterate on a workflow with hot reload, inspect runs in Studio, and verify durability across worker restarts.
- [Deploy to AGNT5 Cloud](https://agnt5.com/docs/get-started/deploy.md): Promote the workflow you ran in your dev session to a managed environment and trigger it remotely.
- [Improve](https://agnt5.com/docs/get-started/improve.md): Close the loop — add an eval, fix the failure, see the diff.
- [Improve with evals](https://agnt5.com/docs/get-started/improve-with-evals.md): Score your investigator agent against a custom scorer, change the prompt, and measure whether the score moved.
- [Install the CLI](https://agnt5.com/docs/get-started/install.md): Install the AGNT5 CLI, verify it, and authenticate with your account.
- [The AGNT5 Loop](https://agnt5.com/docs/get-started/loop.md): Build → Ship → Run → Observe → Improve. The production loop for reliable AI workflows.
- [Where to go next](https://agnt5.com/docs/get-started/next.md): Pick a direction based on what you want to build next.
- [Observe](https://agnt5.com/docs/get-started/observe.md): Inspect traces, retries, state, and failure recovery in Studio.
- [Quickstart: Run your first workflow in AGNT5 Cloud](https://agnt5.com/docs/get-started/quickstart.md): Create a workflow locally, connect it to an AGNT5 Cloud dev environment, invoke it, and inspect the execution trace.
- [Run](https://agnt5.com/docs/get-started/run.md): Invoke your deployed workflow — HTTP, SDK, schedule, or Studio.
- [Run in cloud](https://agnt5.com/docs/get-started/run-in-cloud.md): Promote the workflow from your dev session to a managed AGNT5 environment and trigger it remotely.
- [Ship](https://agnt5.com/docs/get-started/ship.md): Ship your workflow to AGNT5 Cloud — agnt5 deploy, no Dockerfile, no registry.
- [What you just built](https://agnt5.com/docs/get-started/what-you-just-built.md): The mental model behind the workflow you ran — durable steps, journaled state, and the trace as the system of record.
- [Your first agentic workflow](https://agnt5.com/docs/get-started/your-first-workflow.md): Build a support-triage agent that picks tools, calls them, drafts a reply — all under a durable workflow with a full trace.

## improve

- [Improve](https://agnt5.com/docs/improve.md): Capture feedback, build datasets, run evals, compare models, and gate CI.

## run

- [Run](https://agnt5.com/docs/run.md): Crash resume, trace drill-down, search, cost and latency, dashboards, alerts.

## ship

- [Ship](https://agnt5.com/docs/ship.md): Expose, configure, deploy, promote, and roll back production agents.

## Cookbooks

- [Run any AI agent framework on a durable workflow runtime](https://agnt5.com/cookbooks/bring-your-own-agent-framework.md): Wrap LangGraph, OpenAI Agents SDK, Vercel AI SDK, or plain Python in durable AGNT5 steps.
- [Build a customer support agent](https://agnt5.com/cookbooks/customer-support-agent.md): Search docs, draft replies, wait for review, and turn support outcomes into eval cases.
- [Build a data extraction workflow](https://agnt5.com/cookbooks/data-extraction.md): Call tools, force JSON outputs, recover from malformed responses, and inspect every extraction step.
- [Debug and replay a failed AI workflow](https://agnt5.com/cookbooks/debug-production-run.md): Build a support workflow that fails on malformed LLM output, inspect the trace, patch the step, and recover without repeating completed work.
- [Build a deep research agent](https://agnt5.com/cookbooks/deep-research-agent.md): Run long research jobs with search tools, streamed progress, trace inspection, and completion notices.
- [Build a document processing pipeline](https://agnt5.com/cookbooks/document-processing.md): Extract structured fields, validate them, pause for review, and retry failed document steps safely.
- [Build a durable human-approval AI workflow](https://agnt5.com/cookbooks/durable-human-approval-ai-workflow.md): Pause for approval, survive worker restarts, and execute the final side effect exactly once.
- [Build a durable research agent with approval and recovery](https://agnt5.com/cookbooks/durable-research-agent-approval-recovery.md): Checkpoint search, extraction, artifacts, and final human approval across a long-running report workflow.
- [Build a model comparison workflow](https://agnt5.com/cookbooks/model-comparison.md): Run the same case through multiple models, score outputs, and promote the release candidate.
- [Turn a failed production AI run into an eval](https://agnt5.com/cookbooks/production-run-to-eval.md): Capture a bad production run, convert it into an eval case, and compare fixed prompts before release.
- [Build a RAG chatbot with memory](https://agnt5.com/cookbooks/rag-chatbot-memory.md): Retrieve knowledge, preserve user context, isolate tenants, and trace each answer back to evidence.
- [Retry AI workflow steps without duplicate side effects](https://agnt5.com/cookbooks/retry-without-duplicate-side-effects.md): Use idempotency keys and journaled receipts so retries do not duplicate emails, tickets, or payments.
- [Build a webhook triage agent](https://agnt5.com/cookbooks/webhook-triage-agent.md): Receive events, deduplicate delivery attempts, run async triage, and expose a trace per webhook.
- [Debug AI workflows with traces, not scattered logs](https://agnt5.com/cookbooks/workflow-native-observability.md): Compare log-only debugging with workflow-native traces that preserve inputs, outputs, retries, and state.

## API Reference

- [Create a contact](https://agnt5.com/api-reference/create-contact.md) `POST /v1/contacts`: Add a new contact to your contact list in Protocol. You must provide their Protocol username and phone number.
- [Delete a contact](https://agnt5.com/api-reference/delete-contact.md) `DELETE /v1/contacts/{id}`: Delete a contact from your Protocol contact list. Once deleted, it cannot be recovered.
- [Get contacts](https://agnt5.com/api-reference/get-contacts.md) `GET /v1/contacts`: Retrieve your contacts list from Protocol. You can optionally filter and paginate the results.
- [Update a contact](https://agnt5.com/api-reference/update-contact.md) `PATCH /v1/contacts/{id}`: Update an existing contact in your Protocol contact list. You can update any of the contact's attributes.