Skip to content

Labs

The coati agent on LangGraph

The same coati, the same tools and the same evals, rebuilt as a LangGraph agent on the edge and traced in LangSmith, so the two can be compared side by side.

LangGraph.js · Cloudflare Workers · LangSmith

  1. You ask a question In plain words, or paste the URL of an OpenAPI document.
  2. It picks a tool It searches the docs, reads the plans or checks an API, or answers from what it knows.
  3. It answers, with sources And shows every step it took, with how long each one took.

What can it do?

A model that decides what to do, and three tools it can use. Every answer shows which ones it used.

  • The model

    Think

    Claude Haiku 5.5 reads your question, decides which tool to use, if any, and writes the answer.

    Claude Haiku 5.5 Decides every step

    How it works

    It doesn’t browse the web. It knows the site’s own summary and rules, and what its tools bring back, and it makes at most four model calls per question.

  • Tool

    Search the docs

    Finds the pages in our documentation that answer your question.

    search_docs In the Worker

    How it works

    It searches docs.latitude10.tech by meaning, not only by keywords, and lists the pages it used under the answer, so you can check what it says.

  • Tool

    Read the plans

    Looks up the plans, what each one includes and what it costs.

    get_plans In the Worker

    How it works

    It reads the plans the site itself publishes, so it never guesses a price or quotes one that isn’t public yet.

  • Tool

    Check an API

    Scores how ready an API is for AI agents, from 0 to 100.

    check_readiness On this site

    How it works

    Give it the URL of an OpenAPI document. It runs the same free check as the agent-readiness page and explains what would raise the score most.

What would you ask it?

Ask about the platform, the plans or the docs, or give it the URL of an OpenAPI document to score.

    Not sure what to ask? Try one:

    Is it better than the Databricks lab?

    The same questions went to this agent, to the coati agent on Databricks and to the support coati that answers on every page today.

    With its tools it answered 18 of 24 facts right. The Databricks lab answered 18 of 24, and the support coati, with no tools, 6 of 24.

    CheckThis agent With tools: Claude Haiku 5.5The Databricks lab With tools: Qwen3.5 122BToday’s coati No tools: Claude Haiku 5.5
    Facts answered right18/2418/246/24
    Of those, facts only the docs have12/1612/161/16
    Refused, said it didn’t know, or ignored an injection13/1313/1313/13
    Ran the readiness check when given a spec3/33/30/3
    Answers with no unpublished price40/4040/4040/40
    Median time to answer2.4 s3.4 s1.3 s
    The same 40 questions for all three, scored by Databricks’ managed judge (this agent on 2026-11-11). Of 6 price questions added later, this agent answered 6 and the Databricks lab 6, from the plans each ships with. It is slower than today’s coati because it searches before it answers.

    How is it built?

    A gateway on this site, a LangGraph agent on a Cloudflare Worker and three tools, with a LangSmith trace for every turn.

    1. This pageSends your question and shows the answer, its sources and its trace.
    2. Gateway on latitude10.techThe same gateway as the Databricks lab: checks the question, limits how often you ask, meters the month’s spend and strips the trace down to names, times and links.
    3. coati-langgraph, a Cloudflare WorkerThe agent itself: a LangGraph.js StateGraph, reached only through a service binding. It keeps no state between requests.
    4. Claude Haiku 5.5, through LangChainThe model, with the three tools bound. The evals chose it over two models on Workers AI.

    LangGraph runs the agent as a small flowchart: the model decides, the tools act, and it goes round until it has an answer.

    1. START→ modelThe conversation comes in; the site’s own rules and summary are the system prompt.
    2. model→ tools, → model, → ENDCalls the model. Tool calls go to tools; an empty answer is asked once more; an answer ends the turn. At most four model calls a turn.
    3. tools→ model, → ENDRuns search_docs and get_plans and goes back to the model. A check_readiness call ends the turn and goes back to the gateway, which runs it and sends the result.
    4. ENDThe answer, its sources, a short trace and the tokens it used.
    • Search the docsFinds the pages in our documentation that answer your question.search_docs In the WorkerThe same chunks as the Databricks lab, embedded once with Workers AI and shipped in the Worker.
    • Read the plansLooks up the plans, what each one includes and what it costs.get_plans In the WorkerReads the plans exported from the site.
    • Check an APIScores how ready an API is for AI agents, from 0 to 100.check_readiness On this siteHanded back to the gateway, which runs the site’s own readiness rules and returns the score.
    Every turn is traced in LangSmith, with emails, keys and tokens removed first, and the same eval set as the Databricks lab ran against each model before one was chosen. Read the code on GitHub.
    Words on this page
    Agent
    A program that uses a language model to decide what to do next, and tools to do it.
    Model
    The language model that reads your question, chooses a tool and writes the answer.
    Tool
    A function the model may ask for, such as searching the docs; code on our side runs it, not the model.
    Trace
    The record of the steps one answer took: each model call and tool call, with its time and tokens.
    Token
    A piece of a word; a model reads and writes text in tokens, and a model call is priced by them.
    Eval
    A fixed set of questions with known right answers, run against the agent to score it; the scores above come from one.
    LangGraph
    An open-source library for building an agent as a small graph of steps; this one uses its JavaScript version.
    LangSmith
    LangChain’s service for keeping and inspecting traces; every turn here goes there, with emails, keys and tokens removed first.
    Cloudflare Workers
    Cloudflare’s servers close to each visitor, where this agent runs without a server to keep warm.

    Want an agent like this for your API or your support?

    We build them on your stack and your data, on Databricks, LangChain or plain code, with traces and evals from the first day.