Give your agents everything

One day at our monthly team offsite, I witnessed the wildest demo I’ve ever seen. It was our amazing support person, Jesse, showing his ultra customized Claude Code setup that automatically answers all support tickets. He’s still in the loop, vetting each suggested response and action to be taken, but the investigation, issue creation, and suggested reply were all automated. Then he zoomed out and there were eight of these running all at once. The room chuckled as everyone was reminded how much of a beast Jesse is.

Jesse had to put the hard work in to get to that point. Over time hundreds of lines of investigation and practices to follow were embodied in a support skill that would guide Claude to deliver well investigated responses and followup actions. This setup also heavily used the official Mantle MCP server to access our customer data. Just having that data isn’t enough though, what really set his automation apart was giving it access to even more context, like the codebase, logs, GitHub, readonly production databases, etc. Those are the real unlock. Given the powerful models we have today, they can churn through many sources of data much more quickly than humans can, even running for dozens of minutes at a time.

I inadvertently spurred this organizational change for my own benefit at first. Our logging is comprehensive, but it always takes so much time to parse through the logs to figure out where and when an issue is happening. So often I would be waiting for a Grafana logging query to complete, just to then need to pivot to searching for something slightly different. Rinse and repeat and there goes 30 minutes or more on just debugging a single issue. I had a hunch that my own coding agent could be way faster and more effective at finding the right details in the logs. After some setting up of a local MCP server that connected to Grafana, I was cooking. Immediately it was able to cross reference the codebase logging statements with the actual logs in the system. Add in some production database access and it could literally replay any request at any point in time just with this trifecta of code + logs + production database. Some tuning via Claude skills made the coding agent better navigate and filter the logs by using proper indexes, which service to select, etc.

After showing the team the power and time savings you can get from giving your coding agent all this access, my CTO whipped up an MCP server that provided access to all these services and then some: all the production databases, Elasticsearch, Clickhouse, and logging. All one had to do was login to this MCP gateway service via Google OAuth, then add the provided MCP server connection details to your coding agent of choice. It was an instant hit for everyone in Eng, and even the rest of the company jumped aboard.

Having all these extra sources available to our agents not only benefitted our Support folks, it was a big boost in productivity for everyone at Mantle. Engineers could debug and write more performant database queries, automatically fix bugs, and trace down hard to find production issues with all these tools. Sales and marketing benefitted from being able to probe individual customer behaviour even if we didn’t have tracking added to those places, novel customer segmenting could be performed on the raw database data we have, and finding customers to showcase who have adopted a feature better than anyone else for a testimonial is now so accessible.

After a few weeks, this ended up having some immediate effects. Our Support team of one, Jesse, didn’t have a backlog of tickets anymore, and he even got more time back in his day. The number of tough tickets that needed to be escalated to our Support Engineer was drastically reduced, since most investigations would be resolved with a quick response or easy fixes. Engineering was able to really automate their bugfixing and new feature development since they were able to ship higher quality code.

So what are you waiting for? Go give your agents all the access they need. I didn’t mention much about security, PII, secrets, etc. but those are all considerations to think of before giving agents and other roles access. It’s not a silver bullet either, the agents can still lead you to incorrect conclusions, therefore you still need to know what it’s doing.