# How to evaluate your AI agent sessions

> A step-by-step way to measure your Claude Code and Codex history - how many sessions, how much real work, which products - and estimate what it is worth.

Source: https://groundedby.com/guides/how-to-evaluate-your-agent-sessions (updated 2026-10-29)

You have been running coding agents for months. Here is how to find out what that history holds, in about five minutes.

## 1. Find your history

Most agents keep sessions as files on your machine:

| Agent | Where to look |
| --- | --- |
| Claude Code | `~/.claude/projects` |
| Codex CLI | `~/.codex/sessions` |

Other agents keep it elsewhere, often in a database. [Where every agent stores history](/agents).

Check what you have from a terminal:

```bash
ls ~/.claude/projects | wc -l          # projects with Claude Code sessions
find ~/.claude/projects -name '*.jsonl' | wc -l   # session files
du -sh ~/.claude/projects ~/.codex/sessions 2>/dev/null
```

## 2. Keep it from disappearing

Claude Code removes session files older than **30 days** by default. If your history matters to you, raise the limit before you do anything else. In `~/.claude/settings.json`:

```json
{
  "cleanupPeriodDays": 365
}
```

This keeps sessions from now on; it does not bring back deleted ones. [More on Claude Code's storage](/agents/claude-code).

## 3. Count it

Open the [evaluator](/evaluate), pick `~/.claude/projects` or `~/.codex/sessions`, and wait for the card. It runs in the page, and nothing is uploaded. You get:

- **sessions** - files with at least one turn or tool call;
- **active hours** - time between records, with gaps over five minutes dropped;
- **tool calls** - shell commands, file edits and MCP tools;
- **products** - the developer tools your agents worked with;
- **tiers and an estimated range** - see below.

## 4. Read the tiers

The number that matters is not the total. It is how many **deep** sessions you have: 50 or more turns, an hour or more of active work, with tools in the loop. Those are what buyers ask for. [What makes a session valuable](/guides/what-makes-a-session-valuable).

| Tier | Rule | Reported range |
| --- | --- | --- |
| Deep | 50+ turns, 60+ min, tools | $40 - $200 |
| Working | 15+ turns, 15+ min, tools | $5 - $40 |

The range is an estimate from public prices, never an offer. [The valuation method](/docs/valuation-method).

## 5. Sort what is yours

Before you think about selling anything, split your history in two:

- **Yours:** personal projects, open-source work, your own company if you own it.
- **Not yours:** sessions you ran for an employer or a client. Their code and data are usually theirs, whatever your agent's terms say about outputs.

Only the first group could ever be licensed. [Can you sell your AI coding sessions?](/guides/can-you-sell-your-ai-coding-sessions)

## 6. Decide what to do with it

- **Keep it** as a record of your work. Session files are the most detailed log of how you build.
- **Learn from it.** See which tools your agents use most and where they struggle.
- **License it**, one reviewed session at a time, when buyers ask. [Join the waitlist](/#waitlist) to hear when they do.

## Questions

**How long does an evaluation take?**

A few seconds for a few hundred sessions. The evaluator reads files at disk speed, and a large history of several gigabytes takes about a minute.

**What counts as a good result?**

Dozens of deep sessions - 50 or more turns, an hour or more of active work, with tools - matter more than the total. A small history of long, real sessions beats a large one of quick questions.

**My history looks shorter than I expected. Why?**

Claude Code deletes session files older than 30 days by default. Raise cleanupPeriodDays in ~/.claude/settings.json to keep more from now on.