@omarsar0 My version of owning the memory layer: plain files in git. j-skills saves briefs, tickets and lessons per task, and any agent can resume from them. I can also open a repo from another machine, debug its runs and turn the lessons into better skills. https://t.co/C8mdHpOFLz
@michael_chomsky Yes, and don't do review or CI in the session that wrote the code.
Save state to files + git, open a fresh session, and have a different model review only the diff. Run tests locally before pushing, fix every finding in one pass, cap it at 2 rounds.
Small tickets = fewer loops.
@gregisenberg Can't believe I didn't hook my agent to my mailbox earlier. It's the board between anything, the agent, and me. Deals, invoices, events. Cold mail dies in the scan. I only hear what I actually have to act on.
@mvanhorn@mvanhorn Banger article! I also hooked Grok Bot to Gmail so store mail gets checked against the wishlist. I only hear about offers I’d actually want. Inbox stays a watch, not a second job.
Completely agree. Open models are becoming foundational infrastructure. But open models alone aren’t enough. The world also needs open-source tools and frameworks that make it easy for everyone to build, orchestrate, and deploy AI with them. Innovation shouldn’t be limited to a handful of large companies.
@catalinmpit Do you need to capture the whole market? Its a big planet with lots of potential customers...you only need a few to validate and generate revenue. Thats how I overcome this.
Lately I’ve cared less about the model provider and more about matching models to the job. My workflow has stages (brainstorm, spec, plan, execute, review, learn) and each uses the model best suited to that step. Every cycle produces evals, which I use to keep updating those matches as new models ship.
One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.
Agent loops vs agent graphs, World Cup.
A loop is a player: read the game, act, adjust, repeat.
A graph is the team’s tactical system: roles, transitions, set pieces, and responses to changing conditions.
Great players create moments. Great systems help them win consistently.