Loop engineering is having a moment. Most of the examples are task loops: code, test, fix, repeat until done.
But the more interesting question is: what loops make the agent itself better over time? I call it “meta loop”.
The first "meta loop" I built was memory: Making Memory the Moat.
Session context → Structured handoff → Daily logs → Nightly consolidation → Semantic retrieval
A task loop finishes and its value is consumed. A meta loop finishes and its value persists into every future session.
Have been thinking about meta loops and will be sharing more ideas around it.
What meta loops are you building?
‘Signal density’ should be one of the most defining metrics in AI, alongside model intelligence.
Signal density = useful information per token. It matters in both directions.
Output signal density: can the model say the same thing in fewer tokens? Less filler, more substance. Models that win won't just be more capable. They'll be more concise per unit of capability. Every output token is a cost.
Input signal density: can the agent put more relevant context into fewer tokens? Not "give the model everything." Give it exactly what it needs for this turn. Every input token is also a cost.
Intelligence is whether the model can solve the problem. Signal density is how many tokens it costs the system to solve it.
This gives us a way to judge both sides: Models on their output density. Harnesses on their input density.
Thoughts welcome!
94% of skill tokens reduced in your agent harness with zero quality loss.
(That's $150-200/month in tokens saved)
Built Skiller: a dynamic skill retriever that amplifies the signal in your input context and maximizes the use of your tokens.
The model doesn't get smarter. It gets less distracted.
Open-sourced today 👇
Full story (three escalation levels, consent model, production results): https://t.co/BkVNdst7Lb
Open-sourced the pattern: https://t.co/k3Iu4qdUVs
Thoughts and feedback welcome!
196 errors in code in one week → 0 errors for 8 weeks straight. 11 fixes. No human touch.
I run an AI agent daily with increasingly more capabilities: new tools, new integrations, new scheduled jobs. Every addition used to make it more fragile. Timeouts. Corrupted agent sessions. Entirely failed jobs.
So I gave it the Wolverine mode.
A self-healing layer that traces causal chains backward to the root cause. Fix one root, and 20 downstream errors vanish. Not retry logic. Not dashboards. Actual diagnosis and repair.
Now I add capabilities freely without worrying about fragility. The agent heals itself.
This pattern works for any software that runs unattended, not just agents. If you're interested 👇
I'll be sharing more insights, experiments, and original ideas from my harness-building process here.
Memory architecture. Self-healing. Dynamic skill retrieval. Loop control. All grounded in what I actually run daily.
I've been running a personal agent harness daily for 4 months.
The biggest lesson is simple: agentic systems are not about intelligence. They're about continuity.
A model can be intelligent in a moment. A harness makes it useful over time.