1. Step execution (Sonnet): Work for 30 minutes / 100k tokens at the direction of the Orchestrator

2. Review (Opus): Scrutinize the previous step's work for errors, fidelity to the instructions, fix those and record opportunities to improve the agent configuration + tools to reduce errors and token usage (record those to a file).

3. Self-improvement (Opus): Implement the highest impact self-improvement items that don't require user intervention.

Repeat: Until orchestrator session token budget exhausted (set it to 1M or whatever).

The underlying rationale is to keep each step manageable to maximize adherence to instructions and minimize cost (even cached tokens cost something). Prompt tokens are much cheaper than generated, so to the extent Opus mostly reviews rather than drives that saves a lot too. Self-improvement steps are very expensive but the improvements compound, if you're going to run a job for days or weeks it's way more expensive not to do them.

Edit: I do this in Claude Code with the Anthropic models as well as Qwen family models for offline use.

In my experiments I've had some success modeling the work to be done as a DAG of typed artifacts with a combination of code + LLM doing decomposition, transforms, synthesis, and fitness checking to generate the output. It took me a lot of tries to arrive at that formula and it would be cool to have something more general. I also run part of it against local compute because it would be far beyond my budget to do it all on Opus, so something for that would be nice too.

https://app.opencve.io/cve/CVE-2026-28952

There are some fine points to the O(heap size), for example it's clearly unnecessary for the GC to scan objects that do not themselves contain pointers, and work is somewhat proportional to the total number of objects. Combining numerous small objects into manually managed slices, coming up with ways to make the most numerous items pointer-free, etc.

I learned a bit about this when an analytics workload I had ended up with unacceptable pauses (I think over 1 second), Go's GC is more sophisticated now but I think in any GC runtime you have the same considerations to some degree. Some of the best writing at the time was by Gil Tene, one of the principal authors of the C4 concurrent collector at Azul Systems, starting point here:

https://groups.google.com/g/golang-dev/c/GvA0DaCI2BU/m/SmEel...

For the original issue of GC pauses, a narrow change is to move problem data to non-pointer-carrying types, or the bigger hammer of manually managed slices of those types. The second helps with fragmentation too. Some workloads can be split into multiple processes as a direct way to have smaller heaps. If none of those options are enough then off-heap storage lets you do whatever you want.

I do have some complaints about Go, but one of the big ones has been fixed since I last wrote much Go code and it seems like a fine choice for a lot of applications.