AI Agent
An agent that plans, calls tools and reports each step as it works through a task.
- Tool calling
- State machine
- Human approval
Plan
Give the agent a task
Pick a suggestion above or write your own. The agent plans the work, calls one tool at a time and stops for your approval before anything that writes or sends.
Tools available
search_webSearch the public webread_pageFetch a page and extract fieldsread_docRead internal docs and notesquery_crmRead accounts, deals and contactsquery_ticketsRead helpdesk ticketsclassify_textScore sentiment and topicswrite_summaryDraft structured text with citationscreate_docPublish a wiki pagecreate_tasksAssign CRM taskssend_emailSend from your address
Needs your approval before it runs
Run log
Output
The summary, its sources and run stats appear here when the run finishes.
How it works in production
The demo above runs on a script in your browser. This is the architecture it stands in for.
Task input
Approval inbox
Run log
Streamed events
Planner
LLM with tool schemas
State machine
LangGraph
Checkpoints
Resumable runs
search_web
read_doc
query_crm
write_summary
Human approval
Budgets
Steps, tokens, time
Traces
In production the agent is a state machine, not an open loop. The model proposes a plan, and every step is a typed tool call validated against a schema before it runs. Invalid arguments go back to the model with the validation error instead of reaching your systems.
Tools that change something (sending an email, writing to the CRM) pause for approval. Runs are checkpointed, so an approval can arrive hours later and the agent resumes exactly where it stopped.
- Hard limits on steps, tokens and wall-clock time per run
- Every tool call traced with inputs, outputs, latency and cost
- Failed or rejected runs become evaluation cases for the next prompt change
Build something like this
Tell me about the process you want to improve and the systems it touches. I'll come back with questions and a suggested approach.