Pick the cheapest capable model per task for CRM and email work
A user had their OpenClaw agent rewrite its own code to add a throttle module that picks the cheapest capable model for tasks like CRM analysis and drafting outbound emails.
Done withOpenClaw
- What they did
- Koka Sexton runs OpenClaw as a routing layer over OpenAI, Claude and Gemini. He had the agent rewrite its own code to add a manual-throttle module. The agent then picks the cheapest capable model for each job, such as a 30-row CRM analysis versus writing 20 outbound emails, and switches when he tells it to.
- How it went
- The agent now matches model cost to task complexity and acts as a chief of staff, handling follow-ups and drafts. The source gives no cost savings figures.
- Worth knowing
- The throttle is manual: the agent switches models when he tells it to. It sends external messages only after his explicit approval.
Try it yourself with OpenClaw
Add a model-routing module to yourself that picks the cheapest capable model for each task type, such as [task types, e.g. CRM analysis, drafting outbound emails, summarizing]. Use these models and prices: [list of models and costs]. Show me the code changes and a test run with sample tasks before applying them, and finish by logging which model handled each task and the estimated cost.
Source: substack.com · Undated
Five of these in your inbox every morning
The best things people got an AI agent to do, each with the prompt to try it.
More like this
Manage 18 agents from a Notion workspace
Notion used as the control plane for a fleet of 18 OpenClaw agents.
Automate freelance admin with a morning briefing agent
An OpenClaw agent sends a Telegram morning briefing, drafts weekly client status emails sent after one-tap approval, and chases follow-ups, cutting admin from about 9 hours to 20 minutes.
Triage failed GitHub Actions runs and self-heal flaky tests
A QA engineer's OpenClaw agent reads failed CI logs, classifies failures, posts analysis to GitHub and Slack and opens issues; flakiness fell from 8% to 0.8%, though it once wrongly altered an assertion.
- GitHub Copilot
Run a solo company with eight agent departments
A founder set up CEO, CFO, COO, lawyer, accountant, marketing, CTO, and improver agents that share memory and consult each other.
Delegate calendar, inbox and travel booking to a personal work agent
Authors use Cowork or Claude Code to schedule invites, triage email, book flights and hotels, pull Airtable data and scrape exhibitor lists.
- Custom agent
Schedule meetings by email and update the calendar
The poster built an agent that reads emails, checks calendar availability, replies with a meeting time and adds it to the calendar.