Which Predictor? How GEPA Optimizes a Multi-Agent DSPy Program
A multi-agent DSPy program is really several prompts in a trench coat. Here is how GEPA figures out which agent it is improving, and how it hands the better prompt back.
A personal blog about engineering, learning, and creativity.
A multi-agent DSPy program is really several prompts in a trench coat. Here is how GEPA figures out which agent it is improving, and how it hands the better prompt back.
I tried seven ways to make a DSPy program faster and cheaper: three adapters, TOON, a custom adapter I almost wrote, JSON Patch, and finally diffs. Most lost. The real lesson was that DSPy and GEPA made the search cheap enough to run.
Natural language will make AI accessible. Expert language will make it powerful.
How I got Codex, FlowDeck, XCUITest, Loopback, and a local backend to test KIN's voice calendar path end to end.
The least painful setup I found: Chrome owns the auth window, Supabase owns OAuth and PKCE, and a tiny relay page sends the result back.
Why I think DSPy programs need a managed optimization loop, not scattered experiments.
Introversion and diminished social skills can look identical from the outside, but they are not the same thing.
Drawing parallels between Feature Engineering in traditional ML and Data Curation in AI Engineering.
Content Authenticity is missing the catalyst that made HTTPS Universal.
Introducing two Nouns that make working with DSPy clearer.
Getting granular token usage from nested DSPy modules
I failed at GEPA, then learned a few things.
Streaming Named Chunks from Parallel DSPy Modules
Real-Time Tool Call Updates in DSPy with Status Streaming
Type-Safe Tool Contracts for Multi-Module DSPy Programs
My thinking behind this Blog Site.