The Three Agent Stacks I Run Every Day

I’ve been discussing agent harnesses and interfaces with quite a few people lately and thought it would be interesting to walk through the tools I’m using on a daily basis. I run three sets of tools, all for different purposes.

The daily driver: Pi and OpenCode, with Paseo and T3 Code

My daily driver is built around Pi and OpenCode. T3 Code is a new addition, and Paseo is the interface for web, desktop and mobile.

There are 14 specialist profiles from one set of configs: dev, design, devops, marketing, video, music, admin and the rest. The profiles allow me to customize separate instances of Pi and OpenCode. I wrote about this previously.

Paseo and T3 Code have fantastic desktop and mobile apps with a working terminal. I’m not tied to a desk to see what an agent is doing or to dive into the details.

Models get picked per task, and Jev is used to select the appropriate model for subagents and handoffs. OpenCode Go is the workhorse: one subscription, a spread of frontier models, priced so a dozen agents can run all day. Kimi takes the heavy lifting, for planning and the more difficult tasks. Claude is in the mix here too, but it’s rarely used in this setup. I also run Qwen locally as an experiment, though my hardware is modest, so it’s a testbed rather than something I rely on.

The profiles are outfitted with various extensions, plugins, tools, MCPs and CLIs. That includes Camoufox, a stealth browser for sites that block ordinary fetches, read-only Reddit, LinkedIn and X readers, and self-hosted Firecrawl and SearxNG.

I also spent a lot of time instrumenting everything to collect telemetry, which is recorded into DuckDB and funnelled into Prometheus and Grafana. I use a weekly scheduled process to look for insights and potential improvements.

Claude Desktop

Claude Desktop gets a fair bit of use. I use it because:

• A) I really like all the improvements to the harness and interface over the past month.
• B) Opus 5.5 is chef’s kiss.
• C) I need to keep current with Claude to ensure my training is up-to-date.
• D) I will often steal ideas from Claude and build them out for my own harness.

Omnigent and OpenShell: a security experiment

Omnigent and OpenShell are an experiment centred around security and fully autonomous agents. Omnigent is a control plane for sandboxed sessions: policies, sessions and a UI, running as a containerized server. OpenShell is the enforcement boundary underneath it. Real API keys stay in its credential store and the agent sees synthetic placeholders, and policy decides what the sandbox is allowed to touch. Pi and Hermes run as the sandbox’s main processes.

This helps me understand how to build an agent security posture that can scale in an enterprise.

What tools are you using?

What tools are you using at work and for fun? I’m curious to hear about what others are using and what’s been valuable.