Architect Methods

High-level understanding pages for the AgentCanvas architect-search methods.

This is the method collection — one hand-written, logic-level page per architect method. Each page is written to do double duty: help develop the port, and help a reader new to the method understand what it does, without reading the per-skill markdown. The literal execution view — what each skill actually looks like when it runs — lives in the mirror.

Foundation — read first

All methods sit on top of one shared substrate — the files contract. It governs iter-dir layout, parameter resolution, edit whitelist, profile schema, sentinel convention, and the backend wire format. Read this before diving into any specific method.

Files contract — reader's guide

Universal substrate · all methods

What an iter_M directory looks like, how parameters resolve, what you may and may not edit, the sentinel convention, and the backend API bridge — all on one page. Source contract: .claude/commands/architect/_common/files-contract.md.

The methods

adas

ADAS · meta-agent search (Hu et al. 2024)

archive of agents paper + project page

aflow

AFlow · MCTS over workflows (Zhang et al. 2024)

MCTS code-as-workflow

Paper analysis + the aflow port. Monte-Carlo tree search over code-represented workflows; strongest MCTS-branch representative.

myloop

knowledge-distillation orchestrator · v1

THINK→EXP→DISTILL goal-driven termination

Project's own loop. Optimizes understanding against a user-authored goal.md; agent-performance lift is a byproduct. Eight structured working-memory files, no archive.

At a glance

Method Upstream paper Search shape Skill-tree status
adas ADAS — Hu et al. 2024 Meta-agent search — LLM writes new agent code into an archive Reference page here; active skill — mirror/adas-subagent/
aflow AFlow — Zhang et al. 2024 MCTS over code-represented workflows Active skill — mirror/aflow/
myloop Project's own — no upstream paper Hard-paired THINK → EXPERIMENT → DISTILL; eight-file working memory; goal-driven termination Active skill — mirror/myloop/

How to view

# Recommended: served as part of the docs/ dev server
bash docs/run_dev.sh
# then visit http://127.0.0.1:8092/aas/reference/index.html
# (use the top "AAS" tab from any docs page)

# Or open directly
xdg-open docs/pages/aas/reference/index.html

Source of truth

Each method's algorithm.html is a copy of the canonical file under .claude/commands/architect/<variant>/algorithm.html (where the skill folder still ships one). The skill folder version travels with the skill markdown; this site aggregates them for side-by-side reading. If a copy drifts, the skill folder wins.

PageSource of truth
adas/algorithm.html ADAS understanding page + bundled paper. Current port: adas-subagent (see mirror/adas-subagent/).
aflow/algorithm.html .claude/commands/architect/aflow/algorithm.html
myloop/algorithm.html .claude/commands/architect/myloop/algorithm.html