Why do AI 'evolution' experiments miss good tricks found by other branches that never got shared?

This explores why self-improving agent systems that evolve in separate, walled-off lineages miss out on good ideas, and what changes when lineages share. The clearest evidence comes from Group-Evolving Agents Does sharing experience across agents beat isolated evolution?. In the usual tree-style setup, each child agent inherits only from its own parent, so a clever tool fix found on one branch never reaches the others. When the researchers pooled code patches and execution traces across every agent in a generation, performance rose by 14–20 percentage points. The detail that matters most: five of the eight key tool improvements came from a *different* parent than the agent that ended up using them. Useful discoveries tend to show up on one branch and pay off on another, and isolation cuts exactly that connection.

One reason isolation hurts is that a lineage feeding only on its own output runs into the same limits as any closed self-improvement loop. The corpus's broader account of self-improvement Can models reliably improve themselves without external feedback? argues that pure self-improvement stalls through diversity collapse and reward hacking, and that methods that work bring in outside reference points. To any single branch, the other branches act as that outside reference point: they offer variation it could not have produced alone. A related failure happens inside a single model's reasoning. Reasoning models often drop promising solution paths too early and wander instead Why do reasoning models abandon promising solution paths?. Without a shared record, what one attempt learned is simply lost.

What the corpus adds that you might not expect is that sharing failures matters as much as sharing successes. AutoScientists Can decentralized teams outperform central planners in long-running science? beat centralized planners on biomedical tasks under the same experiment budget. It did this with self-organizing teams that kept competing hypotheses alive *and* passed around what didn't work. AutoResearchClaw's pivot-or-refine loop Can experiment failures drive progress instead of stopping it? makes the same point inside one agent: a failure becomes a signal for the next attempt rather than a dead end. An isolated branch that fails keeps that lesson to itself, so the other branches may repeat the mistake.

Sharing doesn't happen on its own, though, and it can get worse as systems grow. Work on scaling agent populations Does scaling agent populations thin mutual observation? finds that as populations grow, each member sees less of what the others are doing and its link to the group weakens. Adding branches can therefore make isolation worse. Some systems answer this with deliberate structure. SkillClaw How can agent systems share learned skills across users? collects trajectories from many users in one place, refines skills from them, and pushes the updates out to everyone. Dream-RSI Can past discoveries train better exploration policies? goes further and treats the whole tree of past discoveries as a cheap simulator for training better exploration strategies. In that view, a branch's history is shared data to replay later, not just a record of its ancestry.

The short answer is that isolated branches don't fail because each one searches badly. They fail because a discovery's value often shows up somewhere other than where it was found, and isolation keeps it from getting there. Classic evolutionary search can rediscover deep ideas from scratch Can evolutionary search discover machine learning algorithms from scratch?, but rediscovering things independently is expensive. A pooled memory lets the population skip work it has already done.