>Make Your Agent Workflows Harness-Agnostic

Jul 24, 2026•3 min read•view on hashnode ↗

Every AI Tinkerers meetup in Seattle opens with a live survey of the room asking what people are building with. It is organized by Actual.ai, and they publish the results on their site.

The first question is which tool you use. Here is every wave so far:

What's your primary AI coding tool right now? Actual.ai Dev Tools Survey Trends, Dec 8 to July 13.

"What's your primary AI coding tool right now?" Actual.ai Dev Tools Survey Trends, Dec 8 to July 13. Source: actual.ai/survey/trends.

What we see here is simple: harnesses go up and down. Cursor led the field at the start and is near zero now, Claude Code climbed to the top and just posted its first real drop, people are not happy with limits and reliability, and Codex jumped into second place in a single wave.

Now the model people reach for most.

Which model do you reach for most? Actual.ai Dev Tools Survey Trends.

"Which model do you reach for most?" Actual.ai Dev Tools Survey Trends, nine waves, Dec 8 to July 13. Source: actual.ai/survey/trends.

Same story for models. Claude and GPT dominate what people reach for, but the lines still move around underneath them, and I always try the frontier Chinese models too rather than assume the leaders stay put.

So go agnostic

The takeaway is simple: make your workflow independent of any single harness or model. Keep the parts that matter outside the tool. Composable skills and a shared memory the agent can reach from anywhere mean that switching harnesses does not leave you starting over.

Across harnesses I move between Pi, Claude Code, and Codex, run my own Hermes instances, and learn each from the CLI on its own terms. For models, alongside my Codex and Claude subscriptions I keep an OpenCode Go subscription at $10/month, which deploys new Chinese models very fast, and $10-20 on OpenRouter to experiment with everything else. For memory I use agentmemory, which the agent recalls from no matter which harness is running.

All of it lives in one repo that any agent can reach. It holds my skills, runbooks that explain to an agent how my environment works, and my custom Pi extensions, of which I add one or two every month to expand what my harness can do. The MCP profiles from my previous post live there too. The repo is agent-agnostic on purpose: whatever harness I open, it points at the same repo and picks up the same setup.

None of it costs me a migration, because the workflow lives outside the tools. Pick tools for the current quarter. Build the workflow to outlast them.