Skip to main content

Documentation Index

Fetch the complete documentation index at: https://internal.september.wtf/llms.txt

Use this file to discover all available pages before exploring further.

The Engine ships with a web_search platform tool that the agent can call to look up information on the open web. It uses Brave Search by default, with Tavily as fallback.

​
When the agent uses it

The agent reaches for web_search when:
  • The question requires recent information (events, prices, releases) the model wouldn’t know from training.
  • The question involves a niche topic or named entity not well-covered by general training.
  • The user explicitly asks the agent to look something up.
The agent doesn’t use it for general knowledge questions. The model answers those directly without burning a search.

​
Configuration

Two environment variables:
NameRequiredDefaultPurpose
BRAVE_API_KEYfor Brave path—Brave Search API key. Free tier: 2000 queries/month.
TAVILY_API_KEYfor fallback—Tavily API key. Free tier: 1000 queries/month.
If both are unset, the web_search tool is registered but returns an error. The agent will surface “I can’t search the web — the search tool isn’t configured.”

​
What the tool returns

A search returns a list of results: 
[
  {
    "title": "...",
    "url": "...",
    "snippet": "...",
    "published_at": "..."
  },
  ...
]
The agent reads the snippets, decides which result is relevant, and optionally follows up by reading the full page (via a separate web_fetch tool, if configured).

​
Quotas

Brave’s free tier gives 2000 queries/month. For an active agent, expect that to last 1–2 weeks. Either:
  • Upgrade Brave’s tier.
  • Cache aggressively at the application layer (same query within an hour returns the cached result).
  • Use Tavily as primary instead of fallback.

​
Pitfalls

  • Searches that should have been memory lookups. If the agent is searching for the user’s own data, that’s a sign retrieval from the brain isn’t surfacing what it should. Look at memory writes during the relevant past sessions.
  • Searches that should have been a known URL fetch. “Find the Anthropic docs” → search → “Read the first result” is wasteful. Add a small docs_search tool that goes to known doc sites directly.
  • Cascading searches. The agent searches, reads the first result, searches again from the result. Five searches deep, you’ve spent significant tokens. Cap the depth in the agent’s system prompt.

​
See also