Improved token efficiency in agentic Sourcegraph tools

We've improved the set of tools available for agents inside Sourcegraph to emit more token-efficient outputs. This includes tools that Deep Search uses under the hood, as well as tools available in our MCP server.

Instead of JSON, the tools now emit a more compact plaintext format, saving tokens on every execution. A lower token count also leads to lower latency and costs. The new format is also more selective about which pieces of information it shows to the agent.

In practice, we've observed a reduction of around 30% in the number of tokens our search tools emit with no loss in output relevancy. This means that agents working with our tools can now run for longer before filling up their context window - whether you're using MCP or Deep Search, you'll be able to get more done.