479-walden-multi-agent-patterns-cognition

topic	agents
type	research
status	research-complete
last-validated	2026-05-21
original-query	Walden Yan multi-agent patterns, single-writer parallel-reader topology, code review loop, Smart Friend cross-frontier routing, ZAO agent squad architecture (reconstructed)
tier	STANDARD

479 — Walden Yan (Cognition) — What's Actually Working in Multi-Agent Systems

Status: Research complete Date: 2026-04-23 Goal: Extract the three proven multi-agent patterns from Walden's April 2026 post and map them onto ZAO OS's agent squad (VAULT/BANKER/DEALER/ROLO/ZOE/ZOEY) + portal bot fleet.

---

Key Decisions / Recommendations

Decision	Recommendation
Default topology for ZAO's multi-agent systems?	USE Walden's rule: writes single-threaded, reads parallel. One writer agent owns the side-effect (tweet, tx, DB write); all other agents contribute intelligence, never actions. This closes the open issue in project_agent_squad_dashboard.md about conflicting agent writes.
Add a Code-Review-Loop to /ship?	USE — a clean-context reviewer catches an average of 2 bugs per PR, 58% severe on Devin. Wire it as a second agent in /ship using no shared context with the coder. This is the single highest-ROI change in this doc.
Use the "Smart Friend" pattern (Sonnet calls Opus on tricky moments)?	USE across-frontier (Claude Sonnet ↔ GPT-5 ↔ Opus) not within-family. Walden explicitly says asymmetric small-to-large is still a training problem; cross-frontier routing works today.
Use manager-Devin-style delegation for long tasks?	USE for 1-week tasks, SKIP for 1-day tasks. Walden's own data: worked for feature batches (10 PRs), broke on short tasks because manager over-prescribes.
Swarm/unstructured networks of agents?	SKIP. "Mostly a distraction." Map-reduce-and-manage is the only shape that scales.

Comparison of Options

Pattern	Walden's verdict	Our fit	ZAO example
Parallel writers (Swarm)	DON'T	Poor	Rejected — matches our own observation in project_openclaw_setup.md
Single writer + parallel reviewers	WORKS	Excellent	/ship + reviewer agent; VAULT tx + BANKER sanity-checker
Small primary calling "smart friend"	WORKS cross-frontier only	Good	ZOEY (Sonnet) delegating hard calls to ZOE (Opus) or GPT-5
Manager + children for 10-PR scopes	WORKS with care	Medium	Portal bot fleet: ZOE concierge dispatching to independent brand bots
Unstructured swarm	DON'T	N/A	—

Walden's Three Patterns In Our Stack

1. Clean-Context Reviewer Loop

Maps to our /review and /ultrareview skills. Action: add a second-agent pass where the reviewer has never seen the implementation chat. Feed it only the diff + the spec. Quantified gain in Devin's numbers: 2 bugs/PR caught, 58% severe (logic, edge cases, security). Attention math: attention-head dilution on long coder contexts is the mechanism.

Implementation sketch:

• src/app/api/ PR created by primary coder (Claude agent with full repo context)
• Reviewer agent spawned with --clean-context + spec + diff only
• Primary agent reviews reviewer findings against user instruction (filters "out-of-scope" and "user explicitly asked for X insecurity")
• Today this is roughly what the santa-method / santa-loop skills already attempt. Upgrade: enforce the no-shared-context rule.

2. Smart Friend (Primary → Larger Model)

Maps to the Opus 4.7 / Sonnet 4.6 split already in our toolbelt (see CLAUDE.md model IDs). Action: the smaller agent decides to escalate on merge conflicts, ambiguous specs, security-adjacent changes. Walden's note: "Primary always calls smart friend at least once to evaluate trickiness" is a cheap prompt-level rule we can adopt immediately.

Walden's also-important point: the smart friend should refuse to guess and instead send the primary to investigate a specific file. Cross-frontier (Claude ↔ GPT-5) works today; small-Sonnet-to-Opus is training-limited.

3. Manager + Children for Multi-PR Scopes

Maps to our "10-order branded bot fleet" plan in project_tomorrow_first_tasks.md: ZOE concierge (manager) dispatches to per-brand bots (children). Walden's warnings we should pre-empt:

• Manager defaults to over-prescriptive — combat by saying "describe the outcome, not the implementation" in the manager prompt.
• Children assume shared state — force an explicit handoff doc (we use ws/ branches per terminal, extend to per-child agents).
• Cross-child messages don't happen by default — add an MCP tool cortex_relationship (see doc 488) so children can leave notes the manager surfaces.

Concrete Integration Points

• src/lib/agents/runner.ts — add a reviewer agent hook after any writer agent completes. Single-threaded write guarantee.
• src/lib/agents/types.ts — extend the AgentRole union with reviewer | manager | child | smart_friend.
• .claude/skills/review.md and .claude/skills/ship.md — add the no-shared-context reviewer pass.
• community.config.ts — no change; this is an agent-infra pattern.
• research/agents/202-multi-agent-orchestration-openclaw-paperclip/ and research/agents/070-subagents-vs-agent-teams/ — cross-reference and supersede the outdated "swarms are cool" sections.

Specific Numbers

• ~8x Devin usage growth in Cognition's largest enterprise segment over the last 6 months.
• 2 bugs per PR caught by Devin Review, 58% severe (logic/edge/security).
• 950 tok/sec SWE-1 (fast sub-frontier) — the model class we'd pair as primary in a Smart Friend setup.
• 200k LOC / 100k LOC / 10k+ iterations — demo-scale multi-agent systems that worked because they had verifiable success criteria. Real product work does not.

What to Skip

• SKIP any design that has 3+ agents taking simultaneous write actions. Every demo we've seen of this pattern fragments decisions.
• SKIP "agent council of equals" framings — the winning shape is 1 writer + N read-only contributors.
• SKIP the asymmetric small-to-large Smart Friend until frontier models are trained for it (Walden flags as open problem).

Sources

• Walden Yan — Multi-Agents: What's Actually Working
• Walden Yan tweet (primary)
• Devin / Cognition
• Anthropic — multi-agent research system blogpost