Internal reference

GP Operators

A visual guide to how AutoAW's genetic programming loop initialises and evolves agent configurations. All animations run live — each card loops through the before → after transition autonomously.

Phase 0 — Seed

Initial population

The GP loop loads one canonical gene per topology type from a fixture file. Each seed receives random temperature jitter across its agents before generation 0 begins.

fixture 01

Fixed Pipeline

researcher → writer

seed2 agents

fixture 02

AI Orchestrated

conditional dispatch

seed3 agents

fixture 03

Debate

advocate + critic → judge

seed3 agents

fixture 04

Parallel Reduce

specialists → synthesize

seed3 agents

fixture 05

Human-in-Loop

human review mid-chain

seed2 agents

fixture 06

Hybrid

broadcast + reduce

seed4 agents

Operators — Mutation

Mutation operators

mutate_structure

randomly selects one of four structural changes to apply

mutation

add_agent

before

remove_agent

before

swap_topology

fixed_pipeline

rewire_edge

before

mutate_prompt

GPT-4o-mini rewrites one agent's system_prompt with a diversity directive

mutation

writer

⟳ LLM rewrite

“Write clear, concise summaries based on research provided to you. Focus on key insights.”

“Craft comprehensive, structured analyses that highlight key insights, implications, and actionable conclusions.”

original system_prompt

mutate_param

Gaussian perturbation on one agent's temperature (σ = 0.1, clamped [0, 1])

mutation

writeragent.temperature

temperature0.42

0.0 · deterministiccreative · 1.0

Δ ~ 𝒩(0, 0.1) → 0.42

before perturbation

Operators — Task-Graph Mutation

Task-graph mutations

Structural rewrites of the task DAG itself. Each mutation has a defined cardinality (input tasks → output tasks or agents) and a mutable attribute — n, k, or target node — that the GP loop treats as a gene.

mutate_expand

1 coarse task → n specialised subtasks; n is the mutable attribute (not hydration)

mutation

1 coarse analyst task

mutate_delegate

n tasks → k agents (0 ≤ k ≤ n); k is the mutable attribute

mutation

k = 3 (1:1)→k = 2 (consolidated)

mutate_inject_critique

insert a critic node after a target executor — 1 node becomes an (executor, critic) pair

mutation

researcher → writer

mutate_compact

merge n adjacent agents into m generalised agents — inverse of mutate_expand

mutation

3 sequential specialised agents

Operators — Crossover

Crossover operators

crossover_subgraph

swaps agent tails between two genes at random split points

crossover

gene Aseed_0001

gene Bseed_0003

split after writer / drafter — tails exchanged between genes

two parent genes

crossover_prompt

swaps system_prompt between agents sharing the same role (50 % probability per shared role)

crossover

gene Awriter

WRIT

system_prompt

“Write concise technical summaries. Prioritize accuracy over style.”

“Craft engaging narratives that explain complex topics to a general audience.”

gene Bwriter

WRIT

system_prompt

“Craft engaging narratives that explain complex topics to a general audience.”

“Write concise technical summaries. Prioritize accuracy over style.”

both writer agents retain original prompts