After shipping 6+ agentic systems with Cursor + Claude, one pattern is crystal clear:
Most projects don't die from bad prompts, they die from missing structure.
Layered harnesses fix this:
• Core project rules
• Per-agent behaviors
• Dynamic shared context
Context drift drops dramatically. Agents stay coherent much longer.
What's the biggest structural challenge you've run into with long-running agents?
I wanted to get familiar with ClaudeDesign and I don't think I have been using my time effectively so decided to build an extension that can track and monitor my time across projects.
I will be releasing this for free as well so others can take advantage of it.
I've been building a small tool to generate cursor harnesses
Nothing fancy
It just helps structure rules and agents better. I'm using it on a few projects and it’s helping a lot
If you are an engineer in the AI space I highly recommend reviewing how Anthropic architected ClaudeCode
You can apply many of those principles into modern agentic systems. Context is key and you can see that based on how they developed the harness for it.
I have seen a bunch of "reply to this" in my feed. I' wondering if bots are sourcing posts for this and just replying.
I guess this post will tell me that 🤷
@markoilico Thank you for this!
I think most of the AI tooling related to UI design lack the creativeness and are extremely repetitive.
Inspiration is really key to empowering them to build a fun experience for users. I could definitely see myself using some of what you built.
Harness is a buzzword rn in AI and I've been experimenting with what a cursor harness actually looks like
- tech specific rules
- agent specific rules
- project level rules
- shared context
Once you start structuring this stuff things feel way more stable.
@aakashgupta Yeah I think you could easily build that orchestration layer with reasoning and understand to what persona applies where etc. Great point though I think it really showcases the need in the industry.
I'm noticing a pattern across Cursor projects.
Rules scattered everywhere
No structure
Agents drifting over time
Feels like most projects are missing a proper "harness"
Especially the more legacy ones.