Clef turns agent routing into a decision-model problem. Set the policy before the model learns it.

Cloudflare introduced Clef and Clef-flash on 1 October as open-source decision models for high-speed classification and agentic workflows, alongside a reinforcement-learning platform for fine-tuning them on a developer's own data. The useful shift is architectural: not every agent decision needs another general-purpose generation call.

That can cut latency and make routing more specialised, but it also moves policy into a model that can be tuned. Once a learned classifier decides which tool, workflow or model receives a request, its errors become control-plane errors rather than ordinary prediction mistakes.

Define the decision boundary before fine-tuning

Start by separating advisory decisions from authority-bearing ones. A model choosing a document category is different from one choosing whether an agent may invoke a production tool. For the second class, define the allowed actions, deny conditions and escalation path before collecting reinforcement signals.

Training reward should not become the policy specification. A reward function can optimise observed outcomes while missing constraints that matter only in rare or adversarial cases. Keep hard authorisation, identity and environment restrictions outside the learned decision model, then let the model optimise only inside that permitted envelope.

A faster decision model is useful only when the system already knows which decisions it is allowed to make.

Evaluate routing errors by consequence

Classification accuracy alone is too coarse for agent routing. Build an evaluation set around consequential boundaries: safe versus privileged tools, low-cost versus high-capability models, internal versus external data paths and automatic versus human-approved actions. Weight false permits more heavily where a wrong route can change systems or expose data.

Because Cloudflare is positioning Clef for agentic workflows, teams should also test distribution shift. Prompts, tool inventories and business rules change after deployment. Track decision confidence and downstream outcomes, and send uncertain or novel cases to a deterministic fallback rather than silently accepting the learned route.

Treat RL updates as policy releases

Cloudflare's new platform allows developers to fine-tune decision models with their own data. That makes the training dataset, reward design and resulting checkpoint part of the production control surface. Version them together, record which policy generated the training signal and require regression tests before promoting a new checkpoint.

The operational advantage is clear: specialised decision models can make agent orchestration cheaper and faster. The governance requirement is equally clear. Keep hard permissions deterministic, evaluate errors by consequence and release every tuned decision model as a policy change, not merely a model upgrade.

Written by the Devence Lab research team.