effect-first research
This site is the active research front-end for a simple question: what guidance actually helps an agent write stronger Effect code, and what turns out to be noise?
effect-first research
This repo is a research effort about one question: how do we build agents that produce better Effect code, more reliably, with less wasted context?
What this project is trying to prove
The long-term goal is definitive, citable evidence about what actually helps an agent write strong Effect-first programs.
That means testing guidance mechanisms directly, keeping the raw artifacts, and being clear about what the evidence does and does not support.
What changed
Earlier versions of this repo treated a website as the primary guidance surface. That historical approach is preserved at /old.
The active direction is simpler and more skeptical: test whether local, task-relevant repo context helps more than generic prompting, and only keep ideas that survive measurement.
What is being tested now
The current experiment asks whether a local-first setup can help an agent solve an Effect task better than a baseline prompt.
The first completed run is intentionally small. It is a screening experiment, not a final answer.
What the current result means
The current result is encouraging, but narrow.
It shows an early signal that local repo guidance may help on this task. It does not yet establish a broad claim about how Effect agents should be built in general.
What stronger evidence would require
To make a result that someone could reasonably cite as strong evidence, this project still needs more tasks, more runs, and more than one environment or model path.
Until that exists, the honest framing is: promising research direction, not definitive proof.
Where to look next
- Read the current protocol in experiment/PROTOCOL.md
- Read the raw result in experiment/results.json
- Study the example corpus in examples/
- Compare against the preserved historical site at /old
Context
Current status
Status: complete. Current conclusion: keep-investigating.
Three internal constrained-context trials were captured on March 3, 2026. The local-only arm compiled and passed the fixture test suite in all 3 trials, while the baseline arm compiled but failed the same completion assertion in all 3 trials. This is a narrow keep-investigating signal only, not a proof claim.
What has been shown so far: the local-only arm won 3 out of 3 internal trials on todo-repo-local-only-001.
What this does not show: it is still not broad or definitive proof of how Effect agents should be built in general.
Read the evidence
Current protocol · Raw result · Preserved historical site
Example corpus
The example projects remain the practical local context set for ongoing experiments.
Research goal
The long-term goal is citable evidence about what guidance actually improves agent performance in Effect codebases. The current result is an early step toward that, not the final answer.