Token consumption in agentic development is rising fast: more agents, richer loops, deeper tool use. The instinct to simply “use fewer tokens” often leads teams to cut capability rather than improve efficiency.

This session explores practical approaches for building token-aware agentic systems:
- Progressive disclosure of context: index-first design and on-demand loading of focused skills and instructions
- Separating design-time reasoning (frontier models) from runtime execution (more efficient models) through reusable, versioned artefacts
- Matching model tier to task complexity so cheaper models are used where they add value without creating rework

We will also examine common architectural traps (such as always-on tool descriptions and over-aggressive context stripping) and organisational practices that help teams adopt these patterns consistently: Centres of Excellence, governance of the agent/skill supply chain, and secure distribution mechanisms.

The goal is to treat token efficiency as a design concern from the start, so organisations can scale agentic workflows without losing control of cost or quality.

Preferred duration: 45 min (adaptable 30–60)
- Audience: Engineering leaders, platform/DevEx, FinOps, architects & senior developers working with agentic workflows
- Level: Intermediate–advanced
- Format: Talk + real-world examples / discussion of practices
- Tech needs: Projector + HDMI + Internet Connection