checkpoint

advancedspreading

A checkpoint is a save point in an AI agent's work: if something goes wrong, you can roll back here and try again from this point.

In plain terms

If you have played computer games, you know the save point: before a difficult level, the game saves the state, and if you lose, you continue from there, not from the beginning. The checkpoint is the same in the work of an AI agent. The agent records how things stood before it intervened.

This is necessary because an agent not only answers but also acts: it rewrites files, works in multiple steps, and can make mistakes. The checkpoint ensures the error is not permanent. With a rewind, you return the files and conversation to an earlier state, as if the bad steps never happened.

It differs from the text editor's 'Undo' button because it does not undo a single keystroke but an entire work session of the agent. In many tools, you can choose separately: restore only files, only the conversation, or both.

The term has a second meaning: in long-running agents, checkpoints also serve to resume work where it left off after an interruption (power outage, error, waiting for human approval).

How it works

Before the agent modifies anything, the system automatically saves the current state of the affected files and the conversation's progress so far.
Every new instruction or major step creates a new save point, turning the work into a series of retrievable milestones.
If you do not like the result, you call up the list of points (for example, with a rewind command) and choose which one to return to.
The system restores files, the conversation, or both to the state at the selected moment.
From here, you can issue new instructions, and the agent will proceed via a different path. During long, autonomous runs, this same save point enables continuation after a stop.

An everyday example

You ask the coding agent to redesign your webshop's payment page, but twenty minutes later you see it went in the wrong direction and rewrote three other files. You do not need to manually untangle what changed: you revert to the checkpoint before the instruction and try again with a more precise formulation.

Who is affected

developerYou can allow the agent to make larger transformations more boldly, because work going in the wrong direction can be undone in one step. → Code

entrepreneur, executiveIntroducing an agent carries less risk where erroneous steps can be reversed, so it is worth asking about this when selecting tools.

employeeFor anyone who has an agent work on documents or spreadsheets, a save point is the safety net if the result is worse than the starting point. → Work & HR

A common misconception

MythIf there is a checkpoint, the agent can do anything because everything can be undone anyway.

In realityA checkpoint typically restores only what the agent modified in local files using its own editing tools. Sent emails, deleted records from remote databases, initiated payments, or launched websites cannot be rolled back.

Limits and risks

Operations performed in the outside world (sending messages, making purchases, modifying remote systems) are not affected; human approval is still required for these.
Many tools only track the agent's direct file edits, not necessarily deletions or renames run from the command line.
It does not replace regular backups and version control: the tool keeps checkpoints locally and for a limited time, after which it may automatically delete them.
When rolling back, the good parts completed since then are also lost, so first check what you will discard.

Similar, but not the same

version control (Git) – Version control is a permanent, shareable, consciously maintained log of a project's history, while a checkpoint is a quick, automatic, and temporary safety net within a session.
modell-checkpoint – In machine learning, a checkpoint refers to the model's saved weights during training, not a restoration point for an agent's work.
agent memory – Memory preserves what the agent remembers for later tasks, while a checkpoint records the full state at a specific moment for restoration.

Related concepts

Broader

AI-agent, agent oversight, state management

Narrower

rewind, session continuation, conversation branching

On the edge

coding agent, human approval, version control, sandbox, agent memory

A brief history

Checkpoint is an old computing solution: the state of long calculations and databases has been saved periodically for decades so that after an error, everything does not have to be restarted from scratch. With the spread of AI agents, it entered coding tools and agent-building frameworks in the mid-2020s, along with the rollback function.

Where things stand

As of late 2026, the leading coding agents and several agent-building frameworks offer this as a core feature. However, reversing external actions beyond files remains unsolved.

Status: 2026 November

Also known as

checkpoint · checkpointing · rewind · checkpoint · save point · restore point · rewind · checkpointing

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.