Agent examples

Small test tasks that make an agent’s execution easy to follow. See what it was asked to do, which files it started with, and how it reached an answer.

Recorded model runs

These four tasks cover edits, project instructions, and recovery from a missing file. They run in tiny, temporary workspaces; they do not measure performance on a real codebase.

Each recording includes the original task, starting files, expected result, and independent checks. Step through the turns to inspect the tool calls and final answer.

Recorded with gpt-5.4-mini (high reasoning) on 2026-10-01. Browsing them makes no model calls.

Download the recordings ↓

Fix cents roundingPassed ↗

Task given to the agent

Fix the cents conversion in price.mjs. Investigate the tests, edit the implementation, and verify the fix. Keep unrelated files and existing tests unchanged.

Starting workspace

The agent starts in the project root. price.mjs rounds prices down, so the existing checks fail for values such as 1.005.

What success looks like

cents(1.005) returns 101, cents(12.345) returns 1235, and cents(0) returns 0. The agent runs the checks and preserves the tests and unrelated files.

Read the starting files and project instructions

`AGENTS.md`

Read relevant source before editing. Run node check.mjs from the project root after changes. Do not modify check.mjs or unrelated.txt.

`price.mjs`

export function cents(value) { return Math.floor(value * 100); }

`check.mjs`

import assert from "node:assert/strict";
import { cents } from "./price.mjs";
assert.equal(cents(1.005), 101);
assert.equal(cents(12.345), 1235);
assert.equal(cents(0), 0);
console.log("price checks passed");

Each workspace also contains unrelated.txt with existing user work that must be preserved.

Turn 1 / Tools chosen by the agent
list
Tool input
```
{
  "path": null
}
```

Turn 2 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_v05iiqKe49pTOz9KW4OkXO29",
    "tool": "list",
    "ok": true,
    "value": {
      "path": "<workspace>",
      "entries": [
        {
          "name": ".git",
          "kind": "directory"
        },
        {
          "name": "AGENTS.md",
          "kind": "file"
        },
        {
          "name": "check.mjs",
          "kind": "file"
        },
        {
          "name": "price.mjs",
          "kind": "file"
        },
        {
          "name": "unrelated.txt",
          "kind": "file"
        }
      ],
      "truncated": false
    }
  }
]

read

Tool input

{
  "maxLines": 200,
  "path": "price.mjs",
  "startLine": 1
}

read

Tool input

{
  "maxLines": 400,
  "path": "check.mjs",
  "startLine": 1
}

Turn 3 / Tools chosen by the agent

Results the model received from the previous turn (2)

[
  {
    "id": "call_LgZpjUu3hRUJtmdF4BYzGyDB",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/price.mjs",
      "content": "export function cents(value) { return Math.floor(value * 100); }\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  },
  {
    "id": "call_00nNQ7ijZnBePB71s9qqV6iE",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/check.mjs",
      "content": "import assert from \"node:assert/strict\";\nimport { cents } from \"./price.mjs\";\nassert.equal(cents(1.005), 101);\nassert.equal(cents(12.345), 1235);\nassert.equal(cents(0), 0);\nconsole.log(\"price checks passed\");\n",
      "startLine": 1,
      "endLine": 7,
      "totalLines": 7,
      "truncated": false
    }
  }
]

write

Tool input

{
  "content": "function toPlainString(value) {\n  const text = String(value);\n  if (!text.includes(\"e\")) return text;\n\n  const [coefficient, exponentText] = text.split(\"e\");\n  const exponent = Number(exponentText);\n  const negative = coefficient.startsWith(\"-\");\n  const unsigned = negative ? coefficient.slice(1) : coefficient;\n  const [whole, fraction = \"\"] = unsigned.split(\".\");\n  const digits = whole + fraction;\n  const decimalIndex = whole.length + exponent;\n\n  let plain;\n  if (decimalIndex <= 0) {\n    plain = `0.${\"0\".repeat(-decimalIndex)}${digits}`;\n  } else if (decimalIndex >= digits.length) {\n    plain = `${digits}${\"0\".repeat(decimalIndex - digits.length)}`;\n  } else {\n    plain = `${digits.slice(0, decimalIndex)}.${digits.slice(decimalIndex)}`;\n  }\n\n  return negative ? `-${plain}` : plain;\n}\n\nexport function cents(value) {\n  if (!Number.isFinite(value)) return value;\n\n  const plain = toPlainString(Math.abs(value));\n  const [whole, fraction = \"\"] = plain.split(\".\");\n  const centsDigits = (fraction + \"00\").slice(0, 3);\n  let cents = BigInt(whole) * 100n + BigInt(centsDigits.slice(0, 2));\n\n  if (centsDigits[2] >= \"5\") cents += 1n;\n\n  const result = Number(cents);\n  return value < 0 ? -result : result;\n}\n",
  "path": "price.mjs"
}

Turn 4 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_xDBkU9hEaU4kkz62aMvP0Qqg",
    "tool": "write",
    "ok": true,
    "value": {
      "path": "<workspace>/price.mjs",
      "bytes": 1214
    }
  }
]

bash

Tool input

{
  "command": "node check.mjs",
  "cwd": "<workspace>",
  "timeoutMs": 30000
}

Turn 5 / Agent’s final answer
Results the model received from the previous turn (1)
```
[
  {
    "id": "call_yXCoGjFqyswTNmdBlkkw487j",
    "tool": "bash",
    "ok": true,
    "value": {
      "cwd": "<workspace>",
      "stdout": "price checks passed\n",
      "stderr": "",
      "exitCode": 0,
      "signal": null,
      "timedOut": false,
      "truncated": false
    }
  }
]
```
Fixed `price.mjs` so `cents()` now rounds monetary values correctly instead of flooring them, handling cases like `1.005 -> 101` and `12.345 -> 1235`. I verified the change by running `node check.mjs`, and it passed (`price checks passed`).

Independent check: price checks passed · 0 compactions

Implement a greeting from project instructionsPassed ↗

Task given to the agent

Implement greet(name) in src/greet.mjs using the project instructions already supplied. Read and edit only greet.mjs, then verify by running exactly node ../check.mjs. Do not reread AGENTS.md or inspect the test file. Preserve existing tests and unrelated files.

Starting workspace

The agent starts in src/. greet() always returns “hello”. Root and local AGENTS.md instructions must be present in its first model request. The local instructions specify the greeting format.

What success looks like

Using the supplied instructions, the agent reads and edits only greet.mjs, then runs node ../check.mjs successfully on the first attempt. It must not reread guidance or inspect tests. greet(" Ada ") returns "Ahoy, Ada!" and a blank name returns "Ahoy, friend!".

Read the starting files and project instructions

`AGENTS.md`

Read relevant source before editing. Run node check.mjs from the project root after changes. Do not modify check.mjs or unrelated.txt.

`src/AGENTS.md`

Greeting format is exactly "Ahoy, <trimmed name>!". Empty names use "friend". Run node ../check.mjs to verify.

`src/greet.mjs`

export function greet(name) { return "hello"; }

`check.mjs`

import assert from "node:assert/strict";
import { greet } from "./src/greet.mjs";
assert.equal(greet("  Ada  "), "Ahoy, Ada!");
assert.equal(greet("  "), "Ahoy, friend!");
console.log("greeting checks passed");

Each workspace also contains unrelated.txt with existing user work that must be preserved.

Turn 1 / Tools chosen by the agent

read

Tool input

{
  "maxLines": 200,
  "path": "greet.mjs",
  "startLine": 1
}

Turn 2 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_PcqCek3WuiurS7tAHr3pZP4n",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/src/greet.mjs",
      "content": "export function greet(name) { return \"hello\"; }\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

edit

Tool input

{
  "newText": "export function greet(name) {\n  const trimmed = typeof name === \"string\" ? name.trim() : \"\";\n  return `Ahoy, ${trimmed || \"friend\"}!`;\n}\n",
  "oldText": "export function greet(name) { return \"hello\"; }\n",
  "path": "greet.mjs"
}

Turn 3 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_LlqE7SsFzHJ7uUx75vf1Ush5",
    "tool": "edit",
    "ok": true,
    "value": {
      "path": "<workspace>/src/greet.mjs",
      "bytes": 137
    }
  }
]

bash

Tool input

{
  "command": "node ../check.mjs",
  "cwd": "<workspace>/src",
  "timeoutMs": 300000
}

Turn 4 / Agent’s final answer

Results the model received from the previous turn (1)

[
  {
    "id": "call_aOVdWkPGYxzJCcyf2KO0FCqH",
    "tool": "bash",
    "ok": true,
    "value": {
      "cwd": "<workspace>/src",
      "stdout": "greeting checks passed\n",
      "stderr": "",
      "exitCode": 0,
      "signal": null,
      "timedOut": false,
      "truncated": false
    }
  }
]

Implemented `greet(name)` in `src/greet.mjs` to return `Ahoy, <trimmed name>!` and fall back to `Ahoy, friend!` for empty or missing names. Verified with `node ../check.mjs` — greeting checks passed.

Independent check: greeting checks passed · 0 compactions

Read a code using nested instructionsPassed ↗

Task given to the agent

Read src/code.txt first, then report its content according to project conventions. Do not edit any files.

Starting workspace

The agent starts in the project root. The requested file is inside src/, which has its own AGENTS.md. The agent must discover those instructions before reading the code.

What success looks like

The agent reads src/AGENTS.md in an earlier turn than src/code.txt and answers "Scoped code: cobalt". It uses only read, list, and search, leaving every file unchanged.

Read the starting files and project instructions

`AGENTS.md`

Read relevant source before editing. Run node check.mjs from the project root after changes. Do not modify check.mjs or unrelated.txt.

`src/AGENTS.md`

Report codes from this directory with the prefix "Scoped code: ".

`src/code.txt`

cobalt

`check.mjs`

import assert from "node:assert/strict";
import { readFileSync } from "node:fs";
assert.equal(readFileSync("src/code.txt", "utf8"), "cobalt\n");
console.log("guidance checks passed");

Each workspace also contains unrelated.txt with existing user work that must be preserved.

Turn 1 / Tools chosen by the agent

read

Tool input

{
  "maxLines": 200,
  "path": "src/AGENTS.md",
  "startLine": 1
}

Turn 2 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_9cNnnAVopsMO2Uz3J0BcHD3V",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/src/AGENTS.md",
      "content": "Report codes from this directory with the prefix \"Scoped code: \".\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

read

Tool input

{
  "maxLines": 200,
  "path": "src/code.txt",
  "startLine": 1
}

Turn 3 / Agent’s final answer

Results the model received from the previous turn (1)

[
  {
    "id": "call_BhOTktSzd8AQeCGwtdIUdLpF",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/src/code.txt",
      "content": "cobalt\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

Scoped code: cobalt

Independent check: guidance checks passed · 0 compactions

Recover from a missing configuration filePassed ↗

Task given to the agent

Read legacy-config.json first. If it is missing, locate the current JSON configuration. Change retries to 3 without changing other settings. Keep existing tests and unrelated files unchanged, and run the checks.

Starting workspace

The agent starts in the project root. legacy-config.json does not exist. The actual configuration is config/current.json, with retries set to 0.

What success looks like

After the missing-file error, the agent locates the current configuration, changes only retries to 3, and runs the checks. The conversation also compacts during this run.

Read the starting files and project instructions

`AGENTS.md`

Read relevant source before editing. Run node check.mjs from the project root after changes. Do not modify check.mjs or unrelated.txt.

`config/current.json`

{"retries":0,"region":"us-east-1","notes":"Keep this exact note."}

`check.mjs`

import assert from "node:assert/strict";
import { readFileSync } from "node:fs";
const config = JSON.parse(readFileSync("config/current.json", "utf8"));
assert.deepEqual(config, { retries: 3, region: "us-east-1", notes: "Keep this exact note." });
console.log("config checks passed");

Each workspace also contains unrelated.txt with existing user work that must be preserved.

Turn 1 / Tools chosen by the agent

read

Tool input

{
  "maxLines": 200,
  "path": "legacy-config.json",
  "startLine": 1
}

Turn 2 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_yQxNmu89IC7OOOiijBjDXz96",
    "tool": "read",
    "ok": false,
    "error": {
      "code": "ENOENT",
      "message": "ENOENT: no such file or directory, open '<workspace>/legacy-config.json'"
    }
  }
]

bash

Tool input

{
  "command": "find . -name AGENTS.md -o -name '*.json' | sort",
  "cwd": "<workspace>",
  "timeoutMs": 30000
}

Turn 3 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_prXjscbXlb7Qd25FFsRIl6oW",
    "tool": "bash",
    "ok": true,
    "value": {
      "cwd": "<workspace>",
      "stdout": "./AGENTS.md\n./config/current.json\n",
      "stderr": "",
      "exitCode": 0,
      "signal": null,
      "timedOut": false,
      "truncated": false
    }
  }
]

read

Tool input

{
  "maxLines": 200,
  "path": "config/AGENTS.md",
  "startLine": 1
}

Turn 4 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_qwU389xYdv1oLmP5yZG5ZKRk",
    "tool": "read",
    "ok": false,
    "error": {
      "code": "ENOENT",
      "message": "ENOENT: no such file or directory, open '<workspace>/config/AGENTS.md'"
    }
  }
]

read

Tool input

{
  "maxLines": 200,
  "path": "config/current.json",
  "startLine": 1
}

Turn 5 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_N8cpaIcGKGiBFRC3RWQfxRCe",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/config/current.json",
      "content": "{\"retries\":0,\"region\":\"us-east-1\",\"notes\":\"Keep this exact note.\"}\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

edit

Tool input

{
  "newText": "{\"retries\":3,\"region\":\"us-east-1\",\"notes\":\"Keep this exact note.\"}",
  "oldText": "{\"retries\":0,\"region\":\"us-east-1\",\"notes\":\"Keep this exact note.\"}",
  "path": "config/current.json"
}

Turn 6 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_SqGCx4VZyDHen9kZ3Pdli9yI",
    "tool": "edit",
    "ok": true,
    "value": {
      "path": "<workspace>/config/current.json",
      "bytes": 67
    }
  }
]

bash

Tool input

{
  "command": "node check.mjs",
  "cwd": "<workspace>",
  "timeoutMs": 300000
}

Turn 7 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_DcJrl544CE3z94H5BxRHKvYQ",
    "tool": "bash",
    "ok": true,
    "value": {
      "cwd": "<workspace>",
      "stdout": "config checks passed\n",
      "stderr": "",
      "exitCode": 0,
      "signal": null,
      "timedOut": false,
      "truncated": false
    }
  }
]

read

Tool input

{
  "maxLines": 20,
  "path": "config/current.json",
  "startLine": 1
}

Turn 8 / Tools chosen by the agent

Results the model received from the previous turn (1)

[
  {
    "id": "call_TG6Jifpis78ghoeTY2iLyMcW",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/config/current.json",
      "content": "{\"retries\":3,\"region\":\"us-east-1\",\"notes\":\"Keep this exact note.\"}\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

read

Tool input

{
  "maxLines": 20,
  "path": "config/current.json",
  "startLine": 1
}

Turn 9 / Agent’s final answer

Results the model received from the previous turn (1)

[
  {
    "id": "call_nrcVNdoQxKRr9vRG6aVEsgNP",
    "tool": "read",
    "ok": true,
    "value": {
      "path": "<workspace>/config/current.json",
      "content": "{\"retries\":3,\"region\":\"us-east-1\",\"notes\":\"Keep this exact note.\"}\n",
      "startLine": 1,
      "endLine": 2,
      "totalLines": 2,
      "truncated": false
    }
  }
]

Done — `legacy-config.json` was missing, so I updated `config/current.json` instead. I changed `retries` from `0` to `3` and left the other settings unchanged. Checks run: `node check.mjs` ✅ (`config checks passed`) No unrelated files or tests were modified.

Independent check: config checks passed · 8 compactions

To run these tasks yourself, see the evaluation script. Live evaluations make paid model calls.

Subagent delegation

This example gives a parent agent two words, red and missing. It delegates each word to a child agent, then combines their answers.

Each child counts and uppercases its word. The uppercase tool deliberately fails on missing; that child reports the error and finishes. The recorded history connects both children to the parent’s calls.

The decisions are scripted and the tools execute locally. This example demonstrates composition and error handling without calling a model.

Read the example ↗ Download the history ↓