Skip to content

Continue a Bayesian optimization run from a checkpoint

Use this example to give an ExecutionAgent a scientific programming task, then return to the same thread with a follow-up visualization request. The agent must implement and run a Bayesian optimization of the six-hump camel function; the second script reopens its checkpoint and asks for convergence and input-importance plots.

This is an agent-generated workflow, not a fixed optimization implementation. Inspect the code, numerical results, and plots that the model produces before relying on them.

What the files demonstrate

  • bayesian_optimization.py starts the OpenAI-backed run, using workspace_BO/ and thread ID BO_test.
  • bayesian_optimization_continue.py reuses that workspace and thread ID so it can continue from the first run's checkpoint.
  • bayesian_optimization_ollama.py is an independent local-model variant. It does not participate in the two-step checkpoint walkthrough.

Read the ExecutionAgent guide for its code-writing and command-execution behavior, and review checkpointing and sharing for the persistence concepts used here.

Prerequisites

  • uv
  • An OpenAI API key for the checkpoint walkthrough
  • A directory whose generated workspace_BO/ contents you are comfortable reviewing and removing

Run every command from this bayesian_optimization directory.

1. Install the example environment

uv sync
export OPENAI_API_KEY="..."

uv sync
$env:OPENAI_API_KEY = "..."

The scripts use openai:gpt-5.4-mini. To select another endpoint, update the model initialization in both checkpoint scripts and follow the configuration guide. Keep both scripts on the same model configuration when comparing the initial and continued runs.

2. Start the optimization

uv run bayesian_optimization.py

The execution agent receives the optimization objective, writes its chosen implementation under workspace_BO/, runs it, and reports its result. URSA also records state for thread BO_test in that workspace and prints a timing summary.

Before continuing:

  1. Read the generated implementation.
  2. Confirm that it evaluates the standard six-hump camel function on an appropriate bounded domain.
  3. Check that the reported best point and value are supported by saved evaluations rather than prose alone.
  4. Review any commands and dependency installations performed by the agent.

Because an LLM chooses the implementation, exact filenames and optimization libraries can differ between runs.

3. Continue the checkpointed thread

Run the continuation only after the first command completes successfully:

uv run bayesian_optimization_continue.py

The continuation script points to the same workspace_BO/, creates a checkpointer from that workspace, and invokes ExecutionAgent with the same BO_test thread ID. Its prompt asks the agent to use the existing evaluation history to create:

  • a convergence plot with the running minimum; and
  • a second plot highlighting important function inputs.

Confirm that the plots use results from the first run. If they silently create a new optimization history, inspect the checkpoint files and first-run output before trying again.

Optional: run the Ollama variant

The Ollama script is a separate choose-and-run example; it does not resume the OpenAI checkpoint. Install and start Ollama, then pull the model named by the script:

ollama pull gpt-oss:20b
uv run bayesian_optimization_ollama.py

ollama pull gpt-oss:20b
uv run bayesian_optimization_ollama.py

Set set_workspace = True in the script if you want this independent run to write under workspace_BO/. Do not assume it can continue the OpenAI thread: the Ollama variant does not configure the same checkpointer or thread ID.

Adapt the Python workflow

Edit the problem string to change the scientific task. If you want a new checkpoint lineage, change both workspace and thread_id consistently in the initial and continuation scripts. Reusing only one of them can attach the follow-up to the wrong state or make the expected state unavailable.

See the Python getting-started guide for model initialization and direct agent invocation patterns.

Troubleshooting and cleanup

  • If authentication fails, verify the active key or provider with the configuration guide.
  • If continuation cannot find prior state, confirm that the first run completed and that neither script's workspace or thread_id was changed alone.
  • If Ollama cannot find its model, run ollama list and make the script's model string match the locally installed tag.
  • To start the OpenAI walkthrough from scratch, move workspace_BO/ somewhere safe or remove it after confirming that you no longer need its generated code, results, or checkpoints.

These scripts make real LLM calls and allow generated code execution. Runtime, cost, dependencies, and artifacts vary by model response.