Ask HN: Has anyone achieved recursive self-improvement with agentic tools?

Things are time bound by instruction creation - at some point you still need a human to dictate the instructions that the orchestrated agents use. From there I've found that -- (1) derive a goal from the instructions (2) break that goal into tasks (3) order those tasks into a DAG (5) spawn the agents to work via the DAG -- seems to be doing everything I want it to do.

The purely code part you described is a bit of an "extra steps" -you can just... vscode open target repo, "claude what does this do, how does it do it, spec it out for me" then paste into claude code for your repo "okay claude implement this". This sidesteps the security issue, the deadly trifecta, and the accumulation of unused cruft.

Self healing, I try two ways:

1) use a memory tool to store learnings for next iteration (Deja.coey.dev) and have the loop system instructions tell how to use it. One orchestrator, and sequential worker agents who run til their context is full and then hand off to the next run with learnings

2) the agent Shelley on exe can search past convos when promoted too for continuation.

I’ve been doing this with great success just “falling” into the implementation after proper guardrails are placed

You could technically instruct the agent to pilot local ollama and launch minions for "dumb" tasks in parallel, but i don't know if it could break out and modify the file system this way... but then, if it resides say in its own VPS, the damage will be contained.

Your agent will basically never get good at learning. The only ways to get closer to that are 1) fine-tuning (expensive, slow and inaccurate), and 2) reinforcement learning (slow and inaccurate). So you can't just build an agent that automatically, incrementally gets better, without waiting for 10+ years for the process to iterate sufficiently. (ask AI researchers, this has been the case for a long time)

However, you can build an agent that can iterate on one specific problem so much that it becomes amazingly good at it, and then do that on another specific problem, and another, until you have a whole bunch of mini-experts. Then you can use those together. To get better than that... use a new model, new prompting techniques, etc.