SkillC: Compressing Autonomous Agent Capabilities via Symbolic Markdown Compilation under Strict Context-Window Constraints

SkillC: Compressing Autonomous Agent Capabilities via Symbolic Markdown Compilation under Strict Context-Window Constraints As Large Language Model (LLM) agents transition from simple conversational interfaces to complex multi-step autonomous reasoning systems, their operational instructions—termed Skills—have grown exponentially in size and complexity. Loading extensive natural language documentation, procedural constraints, and behavioral policies into an agent's active context window induces severe pathologies: quadratic attention overhead, high token latency, financial cost escalation, and the well-documented "Lost in the Middle" attention degradation (context rot). In this paper, we introduce SkillC (where "C" explicitly designates Compressed—Skill-Compressed, rooted in the Markdown Compressed / MDC framework), a deterministic, symbolic compression and compilation paradigm engineered specifically for autonomous agent capabilities. Unlike probabilistic lossy prompt compression techniques that randomly prune tokens or discard critical syntactical negations, SkillC maps verbose natural-language behavioral prose into a compact, domain-specific symbolic state machine while strictly enforcing a Dual-State Invariant (SKILL.mdo to SKILL.md). Crucially, beyond volumetric token reduction, SkillC serves as a cognitive doubt suppressor: natural-language prompts introduce soft semantic hedging ("please try to", "unless necessary"), causing high epistemic uncertainty and behavioral hesitation in the model's decision logits. SkillC systematically excises conversational fuzziness, converting guidelines into crisp Boolean hard gates that collapse the agent's action probability distribution into deterministic certitude. Through empirical benchmarks conducted across a heterogeneous repository of production agent skills, SkillC achieves an average token reduction of 52.4% to 72.8%, a 100% elimination of epistemic hesitation tokens, deterministic adherence to hard behavioral gates (p < 0.001), and an average inference-latency reduction of 38.6%. Source code and benchmarks repository: https://github.com/jean-terrazzoni/SkillC