# AI use cases for Agentdescent (Birfy/agentdescent)

Repo: [Birfy/agentdescent](https://github.com/Birfy/agentdescent) | Use cases: 4 | Page: https://dedupedai.com/use-cases/birfy-agentdescent/

AgentDescent is a Python framework that runs many agents in parallel to propose edits to a shared skill, prompt, or harness, then merges and commits the best changes.

This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL.

- See a skill evolve in one offline demo run: https://dedupedai.com/use-cases/birfy-agentdescent/see-a-skill-evolve-offline-demo/
- Evolve a prompt against your own labeled examples: https://dedupedai.com/use-cases/birfy-agentdescent/evolve-a-prompt-against-your-own-examples/
- Improve an agent skill from inside Claude Code: https://dedupedai.com/use-cases/birfy-agentdescent/improve-an-agent-skill-from-claude-code/
- Compare published self-evolution algorithms with a dry run: https://dedupedai.com/use-cases/birfy-agentdescent/compare-self-evolution-algorithms-without-a-key/

## How to see a skill evolve in one offline demo run with Agentdescent

Repo: [Birfy/agentdescent](https://github.com/Birfy/agentdescent) | Difficulty: Intermediate | Time: 15 minutes | Basis: From the README

Run the packaged demo to watch an evolution loop learn a rule with no API key.

**Who it is for:** A developer evaluating whether AgentDescent fits their prompt or skill work, who wants proof it runs before adding any keys.

**The problem:** Reading about parallel self-evolving agents does not show how an edit gets proposed, merged, and committed.

**The outcome:** A completed offline evolution that prints a held-out reward and the rule it learned.

### Steps

1. `pip install "agentdescent>=0.7.0"`
2. `agentdescent demo`
3. Read the printed line that starts with `what it learned:` to see the rule the loop committed.
4. Optional: `agentdescent doctor` to see which CLIs and keys this machine has.

### Prompt to paste into your agent

```text
In the AgentDescent repo, run `pip install "agentdescent>=0.7.0"` then `agentdescent demo`. Report the held-out reward and the rule learned from the output. Verify that the output includes a line starting with `what it learned:`.
```

### Caveat

The packaged demo runs offline, but the research comparisons need a clone of the repo and are outside the wheel.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/birfy-agentdescent/see-a-skill-evolve-offline-demo/

## How to evolve a prompt against your own labeled examples with Agentdescent

Repo: [Birfy/agentdescent](https://github.com/Birfy/agentdescent) | Difficulty: Intermediate | Time: 45 minutes | Basis: From the README

Point the evolve call at your question and answer rows to get an improved prompt.

**Who it is for:** An AI engineer with a set of question and answer pairs who needs a better system prompt for a model they already call through an OpenAI-compatible endpoint.

**The problem:** Hand-tuning a prompt against examples is slow, and it is hard to tell which edit actually helped.

**The outcome:** An evolved prompt printed with its held-out reward, scored on examples the loop did not train on.

### Steps

1. `pip install "agentdescent>=0.7.0"`
2. `export OPENAI_BASE_URL=https://api.deepseek.com` (or your own endpoint, such as a local vLLM)
3. `export OPENAI_API_KEY=sk-...` (your key for that endpoint)
4. Write a Python script that builds tasks with `tasks_from(rows, prompt="question", gold="answer")`, then calls `evolve(tasks, scorer("exact"), run=run, propose=reflector(model), strategy=SingleSlot(initial_value="You are a helpful assistant."), rounds=8, n_workers=8, max_concurrency=8, held_out_frac=0.3, patience=3, target_reward=0.98)`. (inferred, check the README)
5. Run the script and read `result.rendered`, `result.final_reward`, and `result.outcomes()`.

### Prompt to paste into your agent

```text
In the AgentDescent repo, build a script that loads <path to your CSV of question and answer rows> into tasks and runs evolve with SingleSlot and exact scoring. Expected output: the evolved prompt, the held-out reward, and the outcomes list. Verify that the held-out reward is reported separately from training results.
```

### Caveat

The README's worked example uses the HotpotQA dataset and a DeepSeek model, and its reported gain is for that setup only.

Risk, credentials: The API key is read from the environment, so keep it out of the script and out of any committed file.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/birfy-agentdescent/evolve-a-prompt-against-your-own-examples/

## How to improve an agent skill from inside Claude Code with Agentdescent

Repo: [Birfy/agentdescent](https://github.com/Birfy/agentdescent) | Difficulty: Intermediate | Time: 30 minutes | Basis: From the README

Ask Claude Code or Codex to evolve a skill, with a quoted cost before any run starts.

**Who it is for:** A developer who maintains a shared skill or prompt for Claude Code or Codex and wants it tuned against examples without leaving the agent.

**The problem:** Iterating on a skill file by hand inside an agent means guessing, rerunning, and losing track of what changed.

**The outcome:** A detached evolution run, status checks from the agent, and an evolved skill that is written back only after you approve.

### Steps

1. `git clone https://github.com/Birfy/agentdescent && cd agentdescent`
2. `bash scripts/setup-hosts.sh`
3. `agentdescent doctor`
4. In the agent, say "improve this prompt against these examples" and point it at your example file.
5. Review the call count it quotes, approve the run, then approve or reject the write-back.

### Prompt to paste into your agent

```text
Clone https://github.com/Birfy/agentdescent, run `bash scripts/setup-hosts.sh`, then ask the agent to improve <path to skill or prompt file> against <path to examples file>. Expected output: a call count, a status report, and a proposed evolved version. Verify that nothing is written back to the target file until the user confirms.
```

### Caveat

The plugin setup is documented from the current main checkout, so the install path may differ from the PyPI release.

Risk, credentials: The run needs your model key in the environment, and the agent setup wires into whichever CLIs you have installed.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/birfy-agentdescent/improve-an-agent-skill-from-claude-code/

## How to compare published self-evolution algorithms with a dry run with Agentdescent

Repo: [Birfy/agentdescent](https://github.com/Birfy/agentdescent) | Difficulty: Advanced | Time: 45 minutes | Basis: From the README

Run a dry-run comparison of the ported algorithms without any API key.

**Who it is for:** A researcher or engineer choosing which self-evolution method to adopt, who needs to see how the ports behave before paying for model calls.

**The problem:** Published methods come with different loops and schedulers, which makes fair comparison difficult.

**The outcome:** A side-by-side run of the example comparison, with each port's behavior visible under the same scheduler.

### Steps

1. `git clone https://github.com/Birfy/agentdescent && cd agentdescent`
2. `pip install -e .`
3. `python -m examples.run_demo`
4. Pass the `--dry-run` flag to a port's example script to run it without an API key. (inferred, check the README)

### Prompt to paste into your agent

```text
Clone the AgentDescent repo, install it with `pip install -e .`, and run `python -m examples.run_demo`. Expected output: the comparison results for the example ports. Verify that the run completes without an API key set.
```

### Caveat

Dry runs show the mechanics of each port but do not reproduce the benchmark scores reported for them.

Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/

URL: https://dedupedai.com/use-cases/birfy-agentdescent/compare-self-evolution-algorithms-without-a-key/