Complementing an AI Agent with deterministic tooling using SWAMP | Dieter's blog | Dieter Plaetinck

In my first experiments with AI agents, I quickly learned that there are certain things you should not use them for. Scripts are still king for deterministic tasks, even one-offs. Instead of asking an agent to do a specific task (e.g. modifying CSV records), I ask it to write a script. So I can inspect the logic, refine it, reuse it, and know that I'll work consistently. It's usually faster and cheaper too. Fast-forward to today, and I'm using a toolkit called Swamp that develops this idea much further. It helps agents build (and re-use) deterministic pieces of code, tracks their inputs and outputs, and connects them into workflows. The scripts do the grunt work; the agent handles only the "reasoning glue". This makes the work faster, cheaper, reusable, and easy to inspect and audit. So far, I've used Swamp to: • collect body-weight data from different apps and publish to a Grafana dashboard • reconstruct a location history from sources such as GPX coordinates in Google Photos • investigate an internet problem by doing raspberry pi experiments and querying Prometheus monitoring data These use cases may sound random, but that's the point! Swamp helps with just about any workflow. I recommend trying this awesome tool. More about it here: https://lnkd.in/dFkh9ims