Day 4: Web Search

I continued adding web search capabilities due to the made-up data issue I faced yesterday. The web search plugin is powered by native search for Anthropic, OpenAI, Perplexity, and xAI models but it is costly. For the current grok-4.1-model I can enable web_search and x_search, which run entirely on xAI's infrastructure.

Prompt:

Research and enable web_search for SimpleAgent with adjusted prompt.

- https://openrouter.ai/docs/guides/features/web-search
- https://docs.x.ai/docs/guides/tools/search-tools

Cloudflare Agents SDK v0.2.24 also just released. This update includes MCP client API improvements and fixes a bug where schedules ↗ meant to fire immediately with this.schedule(0, ...) or this.schedule(new Date(), ...) would not fire. Good to upgrade now.

Read the changelog and apply an upgrade to Cloudflare Agents SDK v0.2.24
https://developers.cloudflare.com/changelog/2025-11-26-agents-resumable-streaming/

Time to test. I tried asking for some latest news and even forwarded a Hacker News link to get a summary:

![[day-4-first-issue.png]]

Cool. Got up-to-date responses now.

Slow Response Improvements

Something to notice is that the response is taking quite long

🔍 router-agent (18.37s) → [simple/research/low] → simple-agent (24.70s)

Checking the AI Gateway Logs gives some more insights:

search_tool seems to be taking long
reasoning mode - I don't think I need that for now for trivial questions
Another request shows that a lot of input tokens are being used: 15k tokens.

![[day-4-ai-gateway-logs.png]]

Tool Context Overhead

While researching this, I found an Anthropic blog post about tool context overhead that was just published a few days ago.

https://www.anthropic.com/engineering/advanced-tool-use

"At Anthropic, we've seen tool definitions consume 134K tokens before optimization."

They break down a typical 5-server MCP setup:

GitHub: 35 tools (~26K tokens)
Slack: 11 tools (~21K tokens)
Sentry: 5 tools (~3K tokens)

That is 55K+ tokens before the conversation even starts.

Their solution is a Tool Search Tool that discovers tools on-demand. I have no idea how to start implementing that because they implemented that solution as Claude internal tools tool_search_tool_*.

I gave this article to Claude Plan mode and asked if it had some ideas. It suggests having a separate agent to handle tool selection and RouterAgent picks the suggested tools, which is good for now.

PLAN: Please analyze this https://www.anthropic.com/engineering/advanced-tool-use and suggest some improvements.

The Current Design

After all the iterations, here is the final multi-agent architecture with token optimization and prompting:

Analyze the current codebase and docs/, show me the final architecture diagram

Telegram/GitHub Webhook
         │
         ▼ [Fire-and-forget]
┌───────────────────────┐
│ Platform Agent DO     │
│  (Telegram/GitHub)    │
│                       │
│  Batch Queue:         │ (500ms window)
│  ├─ activeBatch       │ ← Processing (immutable)
│  └─ pendingBatch      │ ← Collecting new messages
└─────────┬─────────────┘
          │
          ▼
┌───────────────────┐
│  RouterAgent      │
│  Hybrid Classify  │
│  ├─ Pattern (80%) │ → 0 tokens ⚡
│  └─ LLM (20%)     │ → 300 tokens
└─────────┬─────────┘
          │
  ┌───────┼────────┬──────────────┬──────────────┐
  ▼       ▼        ▼              ▼              ▼
┌─────┐ ┌────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│Sim  │ │HITL│ │Orchestr. │ │Duyet     │ │Lead      │
│ple  │ │    │ │Agent     │ │Info      │ │Research  │
│Agent│ │    │ │          │ │Agent     │ │Agent     │
│     │ │    │ │ (Plan+   │ │          │ │          │
│     │ │    │ │ Dispatch)│ │          │ │          │
└─────┘ └────┘ └────┬─────┘ └──────────┘ └──────────┘
                    │
            ┌───────┼────────┐
            ▼       ▼        ▼
        ┌────────┐  ┌─────────┐  ┌────────┐
        │Code    │  │Research │  │GitHub  │
        │Worker  │  │Worker   │  │Worker  │
        └────────┘  └─────────┘  └────────┘

┌───────────────────┐
│  State DO         │ ← Watchdog (30s)
│  (Observability)  │
│  ├─ Sessions      │
│  ├─ Traces        │
│  └─ Metrics       │
└───────────────────┘

Github Webhook

Since Telegram seems to be working fine, I asked Claude to implement a similar Github webhook based on the current architecture. Enter plan mode:

Reading the docs and starting to implement a GitHub webhook the same as the Telegram webhook. 
Webhook listen on @duyetbot placeholder.
"eyes" reaction to the comment instead of sending Telegram typing. 
Show "thinking..." with rotation. 
Show the debug footer message with collapsible HTML support.
Remove "eyes" reaction when finished.

First attempt at the implementation did not work. Back and forth with debugging many times until finally got it working:

![[day-4-github-web-search.png]]

![[day-4-web-search-2.png]]

References

Grok API Documentation - xAI's Grok API reference
Cloudflare Agents SDK v0.2.24 - Changelog for new MCP API
Anthropic Tool Use Best Practices - Tool context optimization