Kopium turns agent replies into partial programs
Kopium is a specialized agent harness whose only tool is a Koru interpreter, letting models emit executable, chained programs instead of ordinary tool calls. Its first experiment shows typed outcomes, per-step accounting, and partial-program failures that stop safely when a reply becomes invalid.
Kopium’s strongest idea is treating the wire format as an executable, typed program rather than a transport for opaque tool calls.
- –A single interpreter tool can express multi-step workflows, bind intermediate values, and compose chains the model was never explicitly trained on.
- –Koru declarations generate the grammar, tool vocabulary, documentation, argument shapes, and legal outcome branches enforced by the runtime.
- –Early measurements are promising but provisional: matched tasks used roughly 25–40% fewer context tokens, while a 50-task sweep reached 47/50.
- –Partial execution makes failures diagnosable: completed prefixes remain accounted for, while invalid tails are refused instead of silently producing misleading output.
- –The wire guarantees structure, not judgment; policy/evaluation mismatches and unresolved hard tasks show that better execution semantics still need better benchmarks.
DISCOVERED
1h ago
2026-10-25
PUBLISHED
2h ago
2026-10-25
RELEVANCE
AUTHOR
korulang