OPEN LINK ↗
// 73d agoOPENSOURCE RELEASE
Autoresearch ported to OpenCode as autonomous experiment skill
A developer has ported Karpathy's autoresearch autonomous experiment loop to OpenCode, implemented as a pure markdown skill. The agent autonomously generates hypotheses, modifies code, runs experiments, evaluates results, and logs state — all triggered via a single slash command inside the OpenCode terminal agent.
// ANALYSIS
Karpathy's autoresearch idea is spreading fast, and seeing it adapted as an OpenCode skill rather than a standalone tool shows how the "agentic loop over code" pattern is becoming infrastructure-level thinking for developers.
- –Original autoresearch runs ML experiments autonomously on one GPU; this port generalizes the loop to any code optimization task inside OpenCode
- –Implementation is lean — pure markdown skill plus a TypeScript context-injection plugin, no MCP server required
- –The included demo is striking: BogoSort optimized from 15.6s to 0.002s (7,802x speedup) across 9 autonomous experiments
- –OpenCode's skills system makes this kind of "agentic workflow as markdown" composable and shareable across projects
- –Very early (28 stars, one day old) but the pattern — autonomous hypothesis/experiment/evaluate loops in your editor — is directionally significant
// TAGS
autoresearch-opencodeagentautomationopen-sourcedevtoolcliai-coding
DISCOVERED
73d ago
2026-03-16
PUBLISHED
73d ago
2026-03-15
RELEVANCE
7/ 10
AUTHOR
dabiggmoe2