PaperClub - Compiling Agentic Workflows

What I worked on

I have been thinking about the idea of “renting cognition”. Read some papers looking at different ways to capture repeated reasoning and compile it into something cheaper and more deterministic.

Title	Authors	Year	Link
Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost	Dennis et al.	2026	https://doi.org/10.48550/arXiv.2605.22502
Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation	Trooskens et al.	2026	https://doi.org/10.48550/arXiv.2604.05150
TraceCompiler: Skill-Guided Mining and Compilation of LLM Agent Traces into Mostly Deterministic Workflows	El Yadouni and Li	2026	https://doi.org/10.48550/arXiv.2608.02680
COLLEAGUE.SKILL: Automated AI Skill Generation via Expert Knowledge Distillation	Zhou et al.	2026	https://doi.org/10.48550/arXiv.2605.31264

What I noticed

• Same underlying idea, different compilation targets: workflow, code, skills, or model weights
• TraceCompiler: mine successful traces, find common steps and dependencies, compile them into a mostly deterministic workflow. Keep the LLM where the path is still unclear
• Compiled AI: planner turns a goal + approved tools into a YAML workflow, generates deterministic code, validates it or retries until success. Basically a coding harness /goal for workflows
• Compiling Agentic Workflows moves the procedure into model weights using synthetic data from a frontier model and full parameter fine-tuning. Huge cost advantage as workflow get’s longer. 3B-8B can support 14-55 workflow steps
• Skills capture knowledge, but they still need an LLM to interpret and apply them at runtime

Aha moment

• The system should get less agentic as it learns
• Reason where there is uncertainty, compile what becomes repeatable
• This is another place where owning the harness matters

What still feels messy

• When is a pattern stable enough to compile?

Next step

• N/A