> For the complete documentation index, see [llms.txt](/llms.txt). > A full single-fetch corpus is available at [llms-full.txt](/llms-full.txt). --- title: How AGNT5 works description: "What AGNT5 does with your code: durable runs made of checkpointed steps, pauses that wait for people, and a recorded trace for every run." last_verified: 2026-10-29 category: Getting Started --- AGNT5 runs your agents and workflows durably and records every step. You write ordinary Python, TypeScript, or Go with the AGNT5 SDK. AGNT5 runs that code, saves each step's result as the step finishes, and keeps a trace of the whole run, so a crash, a restart, or a wait for a person doesn't lose work that already finished. You build in your code, run it on AGNT5, and improve it from the runs AGNT5 records. Each change you ship goes around the same loop. ## The pieces - **Your code** runs in a **worker**: a process that uses the AGNT5 SDK and registers your functions, workflows, and agents. On your machine, [`agnt5 dev`](/docs/build/local-development.md) runs the worker. In production, a [deployment](/docs/run/deploying.md) runs it in a managed AGNT5 environment. - **The AGNT5 runtime** starts runs, hands them to your workers, and stores the result of each finished step. - **Studio, the CLI, the API, and the [AGNT5 MCP server](/docs/integrations/tools/mcp.md)** let you start runs, answer pauses, and see what happened in each one. ## Runs survive crashes A **workflow** is a program made of steps. When a step finishes and the runtime accepts its result, that result becomes a checkpoint. If the worker goes down partway through a run, the run doesn't start over. When it resumes, steps with a checkpoint return their saved result without running again, and the step that was cut off runs again. Because a step that was cut off can run a second time, keep calls with outside effects idempotent, such as charging a card or sending an email. [Workflows](/docs/build/workflows.md) covers steps, durable sleep, and parallel work in detail. ## Runs can wait for people A workflow can pause to ask a person something: approve a plan, pick from a list of options, or type an answer. The pause is durable. The run stays paused until someone answers in Studio or through the API, then continues from the point where it stopped, even if the worker restarted while it waited. [Human in the loop](/docs/build/human-in-the-loop.md) shows how to ask and how to answer. ## The building blocks | Block | What it is | | --- | --- | | [Function](/docs/build/functions.md) | A unit of work that AGNT5 can retry. When a workflow calls it as a step, its result is checkpointed. | | [Workflow](/docs/build/workflows.md) | A durable program that sequences functions, agents, and pauses for people. | | [Agent](/docs/build/agents.md) | An LLM in a loop. It follows your instructions and calls tools until it has an answer. | | [Tool](/docs/build/tools.md) | A typed action an agent can call, such as an API request or a search. | | [Prompt](/docs/build/prompts.md) | Model instructions that are versioned with your code and runnable from it. | ## Every run is recorded Each run leaves a **trace**: every step, model call, and tool call, with its input, output, timing, and cost. Open a run in Studio to see why it did what it did. The same records drive improvement. Turn real runs into a [dataset](/docs/improve/datasets.md), score outputs with [scorers](/docs/improve/scorers.md), and compare a prompt or model change against the current version in an [experiment](/docs/improve/experiments.md) before you ship it. ## Find what's failing [Command Center](/docs/improve/command-center.md) turns failed deployments, failed runs, and failed online evals into issues. For each one, an agent reads the evidence and writes a report with the likely cause and a fix prompt for your coding agent. Repeats of the same failure join one issue, and an issue you resolve reopens if the failure comes back. ## Next steps Try it, or read more about the part you need. - [Set up AGNT5](/docs): Install the CLI with your coding agent or from the terminal. - [Run the Quickstart](/docs/quickstart): Build the Hacker News digest workflow and open its trace. - [Write a workflow](/docs/build/workflows): Steps, durable sleep, parallel work, and the patterns that keep runs safe to replay.