# AgentTM: Time-Travel Debugging for AI Coding Agents

> Source: <https://byteiota.com/agenttm-time-travel-debugging-for-ai-coding-agents/>
> Published: 2026-11-10 18:05:20+00:00

Your AI coding agent just rewrote 23 files. The build is broken, and `git log` shows nothing useful. This is the new debugging hell — and it is exactly what [AgentTM](https://github.com/Jyotishmoy12/AgentTM) (Agent Time Machine) was built to fix. Launched on Product Hunt on November 3 by developer Jyotishmoy Deka, AgentTM is a local, open-source tool that wraps your Codex sessions and records every file change as a replayable, restorable checkpoint — without touching your active workspace.

## What AgentTM Does

When you run a Codex session through AgentTM, it captures sanitized session events into a SQLite ledger and takes SHA-256 content-addressed file snapshots at every operation boundary. The result is a complete, immutable timeline of everything the agent touched.

When something breaks — and it will — you open the local React UI, scroll back to the last known-good state, and restore those files directly into an isolated Git worktree. Your active workspace is never modified. You cannot make things worse by looking at what happened.

```
# Install AgentTM
go install github.com/Jyotishmoy12/AgentTM@latest

# Wrap your Codex session
agenttm run codex "refactor the auth module"

# Open the UI to inspect history
agenttm ui
```

All data lives in a `.agenttm/` directory inside your project. No API keys. No cloud sync. No billing surprises.

## The Problem Is Real and Getting Worse

The numbers from a recent survey make the case bluntly. Developers now spend **16.9 hours per week debugging** versus 9.8 hours writing code — debugging accounts for 42% of the average work week. [According to the survey](https://www.infoq.com/news/2026/10/survey-complex-codebases-agents/), 45.2% of developers say debugging AI-generated code takes longer than writing it themselves, and 93% of engineering organizations report at least one bug that was misdiagnosed because of an AI hallucination. One developer publicly disclosed burning $6,000 in Claude credits overnight from an uncontrolled agent loop.

The issue is structural. AI coding agents modify files non-deterministically, run terminal commands silently, and can break things across a repository in ways that do not map to any git commit. Standard observability tooling was not designed for this failure mode.

## Local-First Is the Differentiator

The existing AI agent observability category — AgentOps, LangSmith, Braintrust, Datadog LLM Observability — is almost entirely cloud-based. All of them require an account, API keys, and data leaving your machine. That is a real concern for teams working with proprietary code, or anyone who has experienced a cloud billing incident.

AgentTM runs entirely locally. Here is how it compares to the alternatives:

| Tool | Local? | Free? | AI Agent Focus? | File Time-Travel? | 
|---|---|---|---|---|
| AgentTM | Yes | Yes | Yes | Yes | 
| [AgentOps](https://www.agentops.ai/) | No | No | Yes | Partial | 
| [LangSmith](https://docs.smith.langchain.com/) | No | Freemium | Yes | No | 
| Git only | Yes | Yes | No | No (commits only) | 

It is the only tool in this category that is simultaneously local, free, and built specifically for AI coding agent session replay.

## Built Fast, Iterating Faster

The technical foundation — Go backend, React UI, Wails desktop framework — runs as a native desktop app on macOS and Windows. Version 0.4.1 shipped on November 9, just six days after the initial Product Hunt launch. That release cadence suggests active development and a maintainer who is responding to feedback quickly.

The tool’s current scope is Codex-focused, but the architecture is generic enough that future support for Claude Code, Copilot Workspace, and other coding agents is a logical next step.

## This Pattern Is the Future

AI agent adoption is accelerating fast. [Uber reports 84% of engineers are now agentic](https://byteiota.com/uber-mcp-gateway-production-scale/). The bottleneck has shifted from writing code to understanding and debugging what agents did. The developer role has quietly become 70% reviewer and reviser.

What AgentTM establishes — checkpoint-based session recording with isolated worktree restoration — will likely become the standard audit trail for agentic workflows. Expect larger tools to build this pattern in. For now, [the open-source version](https://github.com/Jyotishmoy12/AgentTM) is worth pulling today, especially if you are running Codex sessions on any codebase you care about.