| topic | agents |
|---|---|
| type | market-research |
| status | research-complete |
| last-validated | 2026-06-09 |
| superseded-by | |
| related-docs | 830, 829, 835 |
| original-query | can we go through all of the x accounts of the links we have to see other similar posts / x articles - mine the high-signal authors from the inbox for more |
| tier | DEEP |
Goal: Zaal asked to mine the X accounts from the inbox for more articles like the ones he saved. X gated guest-token timelines in 2026 (can't bulk-scrape feeds keyless), so this used semantic search on the high-signal authors + their themes. Three rich clusters surfaced - all directly relevant to ZAO's agent stack.
Keyless X timeline mining is dead: guest-token UserTweets now returns an empty timeline (verified live 2026-06-09 - X returns clear-cache instructions, zero tweets). Single tweets/articles still fetch keyless via FxTwitter. So this round used semantic search (exa) on the authors + their topics rather than feed-scraping. That's the keyless way to "keep searching." For exhaustive timeline mining you'd need an auth cookie or a paid API (getxapi ~$0.001/call).
| # | Decision | Why |
|---|---|---|
| 1 | Add a cost-routing layer to ZAO's Workflow/Hermes: classify each task judgment-heavy vs execution-heavy, route to the right model | The 0xRicker cluster's real lesson (the "$62k -> $129" articles) is NOT "Opus + Kimi" - it's the ROUTING rule: "if a clear machine-scorable rule can be written -> cheap executor; else -> Opus." ZAO runs Opus for everything. A routing layer (Opus plans/reviews, Haiku or Kimi-class executes the parallel bulk) is a 100-400x cost cut on execution-heavy fan-outs, same quality. Maps straight onto the Workflow tool's fan-out + the ZOE $50/day cap. |
| 2 | Audit ZAO's ~50 skills against Thariq's 9 categories + 9 tips; adopt "description = trigger", "gotchas section", and PreToolUse usage-logging immediately | Thariq's "Lessons from Building Claude Code: How We Use Skills" is the authoritative skill-authoring playbook (Anthropic, hundreds of skills in prod). It operationalizes doc 829. Three immediately actionable: (a) every skill description is a when-to-trigger, not a summary; (b) a Gotchas section is the highest-signal content; (c) log skill usage via a PreToolUse hook to find undertriggering skills. |
| 3 | Validate Hermes's clean-context critic - Cognition proves a reviewer with NO shared context is SMARTER, not redundant | Cognition's "Multi-Agents: What's Actually Working" found generator+verifier loops work BEST when the agents share no prior context (clean context dodges context-rot, forces reasoning backward from the diff). Devin Review catches ~2 bugs/PR, 58% severe. ZAO's Hermes (coder+critic) should ensure the critic starts clean, not from the coder's context. |
| 4 | Do NOT build unstructured agent swarms - the practical shape is map-reduce-and-manage with single-threaded writes | Both Cognition (Walden Yan) and the Kimi-swarm crowd converge here: multi-agent works when WRITES stay single-threaded and extra agents contribute INTELLIGENCE not parallel write-actions. Unstructured swarms = "a distraction." ZAO's Workflow tool (pipeline/parallel with a synthesis stage) already follows this; keep it. |
The post Zaal saved (0xRicker, "300 agents built a SaaS," doc 830) is one node in a whole movement. The deeper articles (BestHub + Medium "$62,000 -> $129", zenn.dev Claude x Kimi hybrid):
- Architecture: Opus 4.8 = brain (plan, judgment, quality control); Kimi K2.6 Agent Swarm = hands (300 parallel sub-agents, up to 4,000 steps, SWE-Bench Pro 58.6%, $0.6/M tokens). Kimi's own orchestrator is good at task decomposition but NOT high-level design - so Opus owns "what to build," Kimi owns "build it."
- The actual unlock is the routing layer: classify every task first - judgment-heavy (no clear rubric -> Opus) vs execution-heavy (clear machine-scorable output spec -> cheap executor). "$62k -> $129" came from the routing, not just the model swap.
- Discipline that makes it safe: Opus emits a
spec.md(file paths, validation commands, completion criteria) so the executor doesn't wander; executor works on an isolated branch; Opus reviews against a quality rubric before merge. The BestHub piece ships 15 concrete prompts (plan / classify / rubric / review / assemble / handoff-template / cost-tracking) - a ready-made orchestration kit. - ZAO fit: this is a cost optimization for the Workflow tool + Hermes/ZOE. Today ZAO fans out Opus agents; routing the execution-heavy ones to Haiku (or a Kimi-class model) keeps quality and slashes cost on big drains like doc 832 (the 84-item re-fetch ran 22 Opus-class agents - prime routing candidate).
"Lessons from Building Claude Code: How We Use Skills" (Anthropic, hundreds of skills in prod):
- 9 skill categories: (1) Library/API Reference, (2) Product Verification, (3) Data Fetching/Analysis, (4) Business Process/Team Automation, (5) Code Scaffolding/Templates, (6) Code Quality/Review, (7) CI/CD/Deployment, (8) Runbooks (symptom -> multi-tool investigation -> report), (9) Infrastructure Operations.
- 9 authoring tips: don't state the obvious (push Claude out of defaults - e.g. the frontend skill kills Inter font + purple gradients); Gotchas section is the highest-signal content; use the file system + progressive disclosure (skill = folder, point to
references/*.md); avoid railroading (give goal + constraints, not step-by-step); think through setup (config.json+ AskUserQuestion for missing context); the description field is the model-facing "when to trigger," not a summary; memory via${CLAUDE_PLUGIN_DATA}(survives upgrades); store scripts so Claude composes instead of reconstructing boilerplate; on-demand hooks (/carefulblocks rm -rf/DROP TABLE/force-push;/freezeblocks edits outside a dir). - Distribution: check into
.claude/skills(small teams) or a plugin marketplace (scale); curate before release to avoid redundancy; reference other skills by name to compose them. - Measurement: a PreToolUse hook logs skill usage -> find popular + undertriggering skills.
- ZAO fit: the operating manual for ZAO's skill library. The fetch trio +
/inbox+/zao-researchalready match categories 1/2/3/8. Action: add a Gotchas section to each, rewrite descriptions as triggers, and wire usage-logging.
Cluster C - Cognition's "multi-agent that actually works" (mined from walden_yan) - validates Hermes/ZOE
Walden Yan's "Multi-Agents: What's Actually Working" (2026-04-22, the sequel to "Don't Build Multi-Agents"):
- Core finding: multi-agent works today when writes stay single-threaded and extra agents contribute intelligence, not actions. Parallel writers still fragment decisions (style/edge-cases conflict).
- The code-review loop "so stupid it shouldn't work": generator + verifier where they share NO context beforehand. Clean context makes the reviewer smarter - dodges context-rot (attention math degrades long-context decisions), forces reasoning backward from the diff, openly questions choices. Devin Review: ~2 bugs/PR, 58% severe. The communication bridge back to the coder (filtering bugs against user intent) is what prevents looping.
- "Smart Friend": a smaller primary calls out to a bigger model when stuck. Works well cross-frontier (Claude + GPT as a capability router - some debug better, some test better). The weak-primary -> strong-helper version (the big cost unlock) is still an open training problem.
- Higher-level delegation = map-reduce-and-manage: a manager splits scope, children execute, manager synthesizes. Unstructured swarms are "a distraction."
- ZAO fit: direct validation of Hermes (coder + critic) and ZOE (workers + critics). Two upgrades: (1) make the critic clean-context, (2) the "smart friend" pattern = a ZOE worker on Haiku that consults Opus on hard calls (the cost-routing tie-in to Cluster A).
The three clusters converge on one ZAO move: a routed, single-threaded-write, clean-context-critic agent system, packaged as well-authored skills. ZAO already has the shape (Workflow tool, Hermes, ZOE, skill library). The deltas: add cost-routing (Cluster A), apply the skill playbook (Cluster B), make critics clean-context + add a smart-friend router (Cluster C).
- Doc 830 - the inbox cluster where 0xRicker/Thariq first surfaced
- Doc 829 - Anthropic Skills talk (Cluster B's foundation)
- Doc 835 - "AI proposes, jury checks" (same review-rung pattern)
| Action | Owner | Type | By When |
|---|---|---|---|
| Prototype a cost-routing layer for the Workflow tool: classify execution-heavy stages -> Haiku, judgment -> Opus | @Zaal | Build | Next sprint |
| Audit ZAO's ~50 skills vs Thariq's 9 categories + add Gotchas sections + trigger-style descriptions | @Zaal | Skill hygiene | Pre-July |
| Make Hermes critic clean-context (no shared coder context); add a smart-friend Haiku->Opus consult to ZOE | @Zaal | Agent infra | Ongoing |
| Keep mining: cyrilxbt, shawmakesmagic, shannholmberg, eng_khairallah1 next round (cookie or getxapi if timeline depth needed) | @Zaal | Research | Ad hoc |
- Cognition - Multi-Agents: What's Actually Working
[FULL - exa web_fetch; clean-context review, smart-friend, map-reduce-and-manage, single-threaded writes] - Thariq - Lessons from Building Claude Code: How We Use Skills (GitHub mirror)
[FULL - exa web_fetch; 9 categories + 9 tips + distribution/measurement] - BestHub - Cut Dynamic Workflows cost $62k -> $129
[FULL - exa web_fetch; routing layer + 15 orchestration prompts + cost math] - zenn.dev - Claude Code x Kimi K2.5 hybrid (Opus designs, Kimi implements)
[FULL via exa highlight; spec.md handoff, isolated branch, Claude review before merge] - Cognition - Don't Build Multi-Agents
[FULL via exa highlight; the original context-engineering principles] - Latent.Space - Walden Yan / The Age of Async Agents
[PARTIAL via exa highlight; podcast - brain/machine separation, repo setup, multi-agent maturity] - Thariq pinned writing thread (Rattibha mirror)
[FULL via exa highlight; "skills are the abstraction all agents build on", bash is all you need, file system, prompt caching] - Live test 2026-06-09: guest-token X timeline mining returns empty (verified the keyless-timeline wall)
[FULL - primary; UserTweets returned 0 entries]