Independent reproductions will show that compiling reusable agent skills into executable artifacts substantially reduces token consumption without materially degrading task performance.
state: expiredheat: lowuncertainty: highconvergesscott: lowagent-harness token-efficiency skill-compilationVivek Haldar
What is this?
SkillRT is a proposed compilation-and-runtime system that treats reusable agent skills as code and LLMs as heterogeneous processors, decomposing skills into primitive capabilities for more portable and efficient execution. The supplied search summary claims token reductions of up to 40% without degraded task performance, but the case title claims 94%, and the snippets provide neither experimental details nor independent reproductions. They also do not establish Vivek Haldar’s role or identify the researchers behind SkillRT.
Why it matters to Scott
The proposed skill-compilation approach converges directly with Scott’s demonstration-to-agent compilation work and his argument that executable code paths reduce agent context bloat. However, the supplied material has conflicting 40%/94% claims and no experimental details or independent reproductions, so it is currently only another unverified example of an existing position rather than evidence that should change what he builds or argues.
dev:concept.demonstration-to-agent-compilationip:source.why-code-execution-beats-mcpip:concept.skills-and-workflows
queries asked of Scott's wikis
- compiled agent skills versus prompt-loaded skills
- agent harness token-efficiency architecture
- executable artifacts for reusable agent workflows
- cross-model skill portability and capability profiles
- benchmarks for agent token cost versus task performance
- skills as code in coding-agent harnesses
Measured heat
no measured readings yet — the hourly heat pass fills this in
How the heat travelled
no chain yet — the hourly chain pass fills this in
Evidence (3) — ⭐ canonical anchor
Interpretation history
2026-07-27T06:21:16Z
No independent reproduction or task-quality evidence arrived within the observation window; only negligible, increasingly negative engagement remains. The broader selective-context pattern may persist, but this episode has faded without corroborating the specific compilation claim.
2026-07-22T17:30:01Z
The latest activity adds no independent measurements or reproduction of compiled executable skills; the discussion is low-signal amplification of the already-known selective-context result. The specific claim of substantial savings without task degradation remains uncorroborated.
2026-07-22T16:27:23Z
Ratel adds an independent implementation-level result showing large token savings from selectively exposing relevant tools and skills, making the broader context-reduction pattern more credible. It does not reproduce compilation into executable artifacts or adequately establish unchanged task performance, so the specific hypothesis remains uncorroborated.
2026-07-22T16:22:39Z
evidence attached: reddit.post.1v3k4na — Ratel reports a 76% response-token reduction by exposing only contextually relevant tools and skills, directly bearing on compiled or selective agent-skill efficiency.
2026-07-20T18:28:12Z
The added compiler proposal shows adjacent implementation interest but supplies no independent token or task-performance measurements. The core reduction claim remains uncorroborated, with conflicting figures and no evidence that should change Scott’s practice yet.
2026-07-20T18:21:19Z
evidence attached: reddit.post.1v1s4ee — The proposed skill compiler directly bears on whether executable harnesses can replace prompt-only skills, though it provides no independent performance results.
2026-07-20T04:35:48Z
grounded: converges/low — The proposed skill-compilation approach converges directly with Scott’s demonstration-to-agent compilation work and his argument that executable code paths redu
2026-07-20T01:36:10Z
case created — The reported 94% token reduction is a concrete and reproducible harness optimization directly relevant to agent builders.
Decision trace
- 07-27 16:21expireNo independent reproduction or task-quality evidence arrived within the observation window; only negligible, increasingly negative engagement remains. The broader selective-context pattern may persist
- 07-23 03:30repriceThe latest activity adds no independent measurements or reproduction of compiled executable skills; the discussion is low-signal amplification of the already-known selective-context result. The specif
- 07-23 03:20mark_dirtyengagement_update
- 07-23 03:20mark_dirtycomment_update
- 07-23 02:27repriceRatel adds an independent implementation-level result showing large token savings from selectively exposing relevant tools and skills, making the broader context-reduction pattern more credible. It do
- 07-23 02:22attachRatel reports a 76% response-token reduction by exposing only contextually relevant tools and skills, directly bearing on compiled or selective agent-skill efficiency.
- 07-23 02:20propose_attachRatel reports a 76% response-token reduction by exposing only contextually relevant tools and skills, directly bearing on compiled or selective agent-skill efficiency.
- 07-21 04:28repriceThe added compiler proposal shows adjacent implementation interest but supplies no independent token or task-performance measurements. The core reduction claim remains uncorroborated, with conflicting
- 07-21 04:21attachThe proposed skill compiler directly bears on whether executable harnesses can replace prompt-only skills, though it provides no independent performance results.
- 07-21 04:20propose_attachThe proposed skill compiler directly bears on whether executable harnesses can replace prompt-only skills, though it provides no independent performance results.
- 07-20 14:35groundThe proposed skill-compilation approach converges directly with Scott’s demonstration-to-agent compilation work and his argument that executable code paths reduce agent context bloat. However, the sup
- 07-20 11:36createThe reported 94% token reduction is a concrete and reproducible harness optimization directly relevant to agent builders.