Early benchmarks: 135/135 trace-seeded replay cells matched across three model families
Explore the spec Try the playground

Press release · For immediate release

Doug Fir Labs Launches Scholialang, an Open Protocol for Portable AI Agent Reasoning

Scholialang v0.6 introduces content-addressed reasoning traces, DAG-backed reuse, and lazy context preludes for AI agents that need auditable work products across sessions, models, and toolchains.

Seattle, WA — June 11, 2026 — Doug Fir Labs today announced Scholialang v0.6, an open structured reasoning language for AI agent workflows. Scholialang gives agents a compact public vocabulary for the reasoning artifacts teams already need to inspect: goals, observations, evidence, findings, decisions, actions, handoffs, retractions, and conclusions.

Unlike ordinary chat transcripts, Scholialang traces are designed to be validated, diffed, stored, referenced, and reused. The v0.6 release adds a content-addressed substrate: optional SHA-256 canonical IDs on atoms, a canonical-ID-keyed DAG registry, canonical-ID-aware references, and three deterministic lazy-prelude modes for carrying prior reasoning forward without replaying entire transcripts.

“Agent work is becoming operational record, but transcripts are not infrastructure,” said Darren Brewster, Founder at Doug Fir Labs. “Scholialang is our attempt to make the public reasoning state of agent work portable: readable by humans, checkable by tools, and reusable by later agents.”

Early pilots suggest the approach can preserve decision-relevant state across model families while reducing carried context. In a three-model reasoning-replay smoke test across Opus 4.8, Fable 5, and GPT-5.5/Codex, Scholia trace carryover matched the original decision in 135/135 trace-seeded cells. In separate compounding pilots, a 150-cell Claude Fable 5 run showed hash_list reducing Session-5 input tokens by 41.0% with no measured quality loss, while hash_only_lazy reached 49.5% reduction with a small quality tradeoff. A Codex full-stack pilot independently reproduced a 30.45% lazy-hash compression result with quality parity under the tested harness. Doug Fir Labs characterizes these as early pilot results, not a final benchmark.

Scholialang is designed to complement existing agent infrastructure rather than replace it. Retrieval systems can fetch prior atoms, validators can reject malformed references, review tools can inspect dependency neighborhoods, and model sessions can resume from compact public reasoning state instead of rediscovering prior work from raw chat logs.

The launch includes the v0.6 specification, Python reference package, MCP/LSP tooling, host plugins, examples, and public documentation. The specification prose is licensed under CC-BY-4.0; reference implementation and tooling code are available under dual MIT OR Apache-2.0 terms.

Availability

About Doug Fir Labs

Doug Fir Labs builds tools and protocols for reliable AI-assisted software work, with a focus on inspectable agent systems, reusable reasoning artifacts, and human-auditable automation. Find out more at dougfirlabs.com.