A small population can find a good agent prompt and still get stuck. As its candidates converge, an unusual low-scoring branch may disappear before a later mutation can make it useful. The Darwin Gödel Machine (DGM) addresses that failure with evolutionary archive search: keep every tried candidate available for future sampling, rather than breeding only from recent survivors. section 4.3.1 section 4.4.3
DGM follows a vary–select–remember loop, but replaces a small, forgetful population with an archive that never truncates. It does not require a mathematical proof that a self-change will help: it runs the change and measures the result. The original system improved Python harness code; this prompt-focused version mutates agent context instead. section 4.0 section 4.3.1
One iteration makes the archive useful rather than merely large:
- Sample a parent, usually from strong candidates but sometimes from anywhere in the archive. section 4.3.5
- Reflect on the parent’s recorded trajectory, mutate its content, and measure the child. section 4.3.5
- Append the child without deleting older candidates; record its score and parent link so the run remains traceable. section 4.3.5
Recovering an abandoned branch
On a revenue task, a main line of prompts climbs to a good score but stalls. A low-scoring, structurally different branch remains in DGM’s archive. Sampling it later leads to phrasing that excludes both cancelled and returned orders, passing the plateau that the converged population could not cross. The win depends on retaining and revisiting that earlier branch, not on changing model weights. section 4.3.5 section 4.3.1
Listing 4.8 Never discard, and sample by exploit-or-explore (06_revenue_dgm_loop.py)
def sample_parent(archive): #A
ranked = sorted(archive, key=lambda e: e["score"],
reverse=True)
if random.random() < EXPLOIT_P: # usually exploit
return random.choice(ranked[:3]) # a top candidate
return random.choice(archive) # else explore any member
for it in range(iters):
parent = sample_parent(archive) #B
diagnosis = await reflect(parent["content"], parent["traj"])
child = await mutate(parent["content"], diagnosis)
score, _, traj = await measure_full(child)
child_art = parent["art"].mutate(new_content=child, #C
created_by="dgm")
await store.record( #D
Variant(artifact=child_art, parent=parent["art"],
search_method="dgm"),
GapMeasurement(score=score),
)
archive.append({"art": child_art, "content": child,
"score": score, "traj": traj}) #EThe listing shows the two decisive operations: occasionally sample from the whole archive, then keep the measured child and its lineage record. section 4.3.5
That memory is expensive: a growing archive and many iterations mean substantially more model calls than a small-population search. Use it when a reliable score justifies pursuing a hard final improvement across a deceptive low-scoring branch. On easier tasks or noisy signals, the extra search may not pay off. A highest-scoring candidate is not automatically the one deployed; promotion remains a deliberate gate. section 4.3.5 section 4.4.3
