- Author
- Bruno Gonçalves, Data For Science, Inc.
- Published
- 15 July 2026
- Kind
- Technical tutorial blog post with companion code
- Original
- data4sci.com/blog/building-an-advanced-agentic-harness
- Subject
- Harnesses and agent loops
- Themes
- Handing over control, Getting good
What it is
Bruno Gonçalves writes for Data For Science. The post is a tutorial that upgrades a basic agent loop from an earlier post. It aims to show the main parts of a custom harness.
What it says
- The post adds seven production-shaped parts to a basic agent loop.
- The author types each tool's arguments so bad calls fail before running.
- The planner emits a task graph, and independent steps run in parallel.
- The post describes memory in three tiers under a hard size budget.
- The post runs cheap checks first and calls an LLM judge only on survivors.
- Budget pressure above 0.9 makes the harness skip the LLM judge.
- The post admits one demo shows the harness can work, not that it works.
What evidence it gives
The post offers Python code snippets and a companion notebook repository. The example is a small city-comparison task using a mocked fact dictionary. No benchmark or eval results are given.
Interests and limits
Data For Science sells consulting and training on LLM systems. The demo uses mock data and can run on a mock model. The post lists its own caveats, such as in-process memory and trusted tool outputs. Eval results are promised for a future post.