Agent Context Optimization

Dawn optimizes agent context by keeping stable prompt sections cacheable and loading tools only when they are needed.

What This Means

Agent runs can include a lot of context: system instructions, workspace policy, user preferences, skills, memory, connected sources, files, and tool schemas. Sending everything on every turn is slow and expensive.

Dawn reduces that overhead in two ways:

stable prompt sections are structured so compatible providers can reuse cached prompt content;
tools that are not immediately needed can be deferred and found through tool search.

The goal is not to hide capability from the agent. The goal is to avoid filling the initial model request with tools and instructions that are irrelevant to the current task.

Prompt Caching

Prompt caching works best for content that is stable across runs:

Dawn’s base system prompt;
workspace-level guidance;
assigned skills that are preloaded;
stable tool and source summaries;
durable user or workspace context that changes infrequently.

Content that changes on every turn, such as the user’s message, active file references, current time, or event payload, is treated as dynamic context.

When a provider supports prompt caching for the selected model, Dawn structures the request so stable sections can be reused. When the provider or model does not support caching, the agent still runs normally.

Deferred Tools And Native Tool Search

Dawn agents can have access to many tools: memory tools, connected-source tools, file tools, document tools, workflow tools, browser tools, local client tools, and provider-native tools. Most prompts only need a small subset.

For models that support provider-native tool search and have it enabled, Dawn uses the provider’s native search path to discover deferred tools. The initial tool set stays small and high-value, while the agent can still find additional tools when the task requires them.

The initial set is reserved for tools that are broadly useful at the start of a run, such as memory access, provider-native tools, skill loading, intermediate result publishing, Dawn docs search, image generation, and connected-source discovery.

What Users Should Do

Keep agent profiles focused.
Avoid assigning every source to every agent.
Use preloaded skills only for compact instructions that should affect most runs.
Keep long, rarely used instructions as normal skills so Dawn can load them on demand.
Use specialized agents for specialized work instead of one broad agent with every capability.

Efficient context helps cost and latency, but it does not replace good agent design.

Troubleshooting