Faster Agent Context With Prompt Caching and Tool Search

Dawn agents now handle large context more efficiently. Stable parts of the agent prompt can be reused through provider prompt caching, and tools can be searched for on demand instead of being pushed into every model request.

This is not a new button in the UI. It is a practical improvement to the way capable agents behave when they have memory, skills, connected sources, command context, and many available tools. The goal is simple: keep agents useful without making every turn pay the full cost of context the model has already seen or tools it may never need.

Rich Agents Need a Lot of Context

A useful Dawn agent is rarely just a model plus a short prompt. It may have a role, workspace instructions, assigned skills, memory pages, connected sources, channel context, user preferences, file references, and tool descriptions.

That context is what makes the agent useful. A pull request reviewer needs the review rules. A support agent needs approved memory. A research agent needs source boundaries. A workflow agent needs to know what kind of output the workflow expects.

The problem is that context can grow faster than the work in front of the user. If every turn carries every stable instruction and every possible tool description, larger workspaces become more expensive and slower than they need to be.

Prompt Caching Reuses Stable Context

Prompt caching helps with the parts of context that do not change between turns. Dawn can separate stable sections such as core instructions, skills, source descriptions, and reusable prompt blocks from the pieces that are unique to the current request.

When the provider supports caching, the model can reuse the stable prefix instead of treating it as entirely new input every time. For the user, that means long-running threads and repeated agent runs can stay richly grounded without forcing the same context to be paid for from scratch on every turn.

This matters most for agents that do serious work: code review, planning, connected-source exploration, document analysis, and background runs that call several tools before answering.

Tool Search Keeps the Prompt Smaller

The other side is tool search. Dawn agents can have access to many tools: connected-source tools, workspace file tools, document tools, browser or local client tools, workflow tools, and provider-specific tools. Passing every tool schema to every model request is wasteful when most requests only need a few of them.

Tool search gives the agent a way to discover the relevant tools for the task at hand. If the user asks about a pull request, the agent can reach for code-review and source tools. If the user asks for file work, it can find file and document tools. If the user asks for a dashboard, it can find the dashboard-related tools.

The user-facing benefit is that agents can stay broad without every answer dragging the full tool catalogue around.

Better Long Runs, Not Less Capable Agents

This change is not about making agents smaller by removing context. It is about being more selective with how context is used.

A good agent still needs the instructions, memory, sources, and tools that make it reliable. The difference is that Dawn can avoid repeating stable context unnecessarily and avoid presenting irrelevant tools before the model has a reason to consider them.

That is especially important as work moves beyond a single chat turn. Workflows, background tasks, browser automation, large file analysis, and multi-step review runs all benefit from agents that can stay capable without accumulating unnecessary overhead.

Works Across Provider Families

Dawn’s model catalog already spans several providers. Prompt caching and tool handling are not identical across them, so Dawn treats this as part of provider support rather than one generic switch.

For users, the model-picker experience remains the same. Choose the model that fits the agent’s job. Dawn handles the context strategy available for that provider path, and the agent keeps the same prompt, skills, memory, and connected-source assignments.

What Teams Should Notice

The most visible result should be steadier performance as agents become more capable. Larger workspaces can keep richer instructions and more connected capabilities without making every interaction feel heavier than the work requires.

The best agents are still well-scoped. A support agent should not have every engineering source. A code reviewer should not need every personal integration. But once an agent has a real job, Dawn now has better machinery for keeping that job efficient.

Getting Started

There is no separate setup step. Keep configuring agents the normal way: choose the right model, assign the right sources, add the right skills, and keep prompts focused on the agent’s job.

If you manage a workspace with several broad agents, this is a good time to review whether each profile is scoped correctly. Prompt caching and tool search help with efficiency, but clean agent design still matters. The best result comes from both: focused profiles and smarter context handling underneath.