Skip to content

837-x-account-mining-round2

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

837-x-account-mining-round2

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

topic agents
type market-research
status research-complete
last-validated 2026-06-10
superseded-by
related-docs 836, 830, 838, 829
original-query keep mining - round 2: mine the remaining high-signal X authors (shaw, shann, cyril, khairallah, ziwen, prateek, saai, peter pang, whedon, kirill) for more agent/AI content
tier DEEP

837 - X Account Mining Round 2 (10 authors, 40 findings)

Goal: Continue the doc-836 mining across the remaining high-signal inbox authors. Keyless semantic-search method (X timelines stay walled). 13-agent workflow, ~875k tokens. Six clusters - the field is converging hard, and several findings map directly onto ZAO's stack.

Key Decisions

# Decision Why
1 The recurring external pattern IS ZAO's stack - "recursive orchestrator + CLAUDE.md leverage + persistent memory + clean-context review." Stop treating these as new; treat them as validation + a parts catalog Across 10 unrelated builders the same architecture recurs: an orchestrator that is itself an agent, driven by a high-leverage CLAUDE.md, backed by a persistent memory layer, with autonomous review loops. ZAO runs all four (Workflow tool, CLAUDE.md, memory/Bonfire, Hermes critic). The findings are a catalog of refinements to steal, not a gap.
2 Adopt the "Obsidian/typed-folder second brain" framing for ZAO's synthesis layer (doc 838 Stage 2) Recurs in 4 findings (cyril, ziwen, 0xsammy, shann's gBrain at 97.6% R@5 LongMemEval). The pattern: typed folders (people/companies/projects/operations), agent reads-first then writes-back, compounds over 90 days. ZAO's memory dir + Bonfire are this; formalize the typed-folder structure.
3 INVESTIGATE SubQ (sub-quadratic attention, 12M context, ~300x cost cut) as a future cost-routing target alex_whedon's SubQ claims drop-in Claude Code replacement at RULER-128K $8 vs $2,600 on Opus. If real, it changes the doc-836 cost-routing math (cheap long-context executor). Watch; verify the benchmark before betting.
4 Steal ComposioHQ Agent Orchestrator's "reactions" pattern for Hermes Open-source, recursive (built by the agents it orchestrates): CI fails -> auto-spawn fix agent; reviewer comment -> route to session; approved+green -> notify. 8 swappable plugin slots. Hermes's auto-PR pipeline should adopt the reaction-router shape.

Findings - six clusters (40 items)

Orchestrators as Recursive Control Planes

10 items. Pattern: move from serial queues to concurrent orchestrators managing 8-300 parallel agents with explicit coordination, shared memory architectures, and feedback loops that route failures back into session context. Orchestrator itself becomes an intelligent agent, creating recursive improvement where agents build the system that manages agents. This pattern enables 40K-line codebases in 8 days and 99.9% accuracy on 12.5M-line repos.

CLAUDE.md as the Control Signal; Eliza as the Web3-Native Blueprint

7 items. Pattern: framework design is determined by single high-leverage configuration files (CLAUDE.md, character.json in Eliza) that encode agent personality, guardrails, and decision-making logic. Success factors in open-source frameworks like Eliza: modularity, out-of-box readiness (Discord/X/Telegram social loops), trust engines for autonomous finance, and TypeScript ecosystem fit. Vague config produces generic outputs regardless of model quality.

  • @shawmakesmagic - Eliza: A Web3 friendly AI Agent Operating System (ArXiv) - Formal research paper documenting Eliza design principles: web3-first, pluggable modular, simplicity-over-completeness. Compares favorably vs LangGraph, AutoGPT, CAMEL on web3 support, social media integration, multi-agent capability.
  • @shawmakesmagic - Shaw on Eliza Framework Evolution & Marketplace of Trust - Eliza = Shaw's fifth-generation framework. Success factors: out-of-box readiness, TypeScript/web3 fit, solving social loop via X/Discord/Telegram. Introduces Marketplace of Trust where agents build credibility via simulated trading performance.
  • @shawmakesmagic - Shaw's Personal Essays: Eliza Labs vs X, Multi-Agent Boredom - Documents Eliza Labs lawsuit against X (demanded $600k/year API licensing). Explores cringeloops in multi-agent simulations: agents trapped in orchestrated turn-taking rather than emergent behavior. Frames Eliza as path to autonomous agent DAOs.
  • cyrilxbt - The Complete Hermes Agent Setup Guide (Masterclass) - CLAUDE.md identified as single leverage point across all workflows. Vague config produces generic outputs regardless of model quality. Persistent SQLite memory + Markdown skill system + configurable scheduler = 90-day compounding where accumulated decision history becomes irreplaceable.
  • @intuitiveml (Peter Pang, CREAO CTO) - Why Your Company's 'AI-First' Strategy Might Not Be Real AI-First - Distinguishes 'AI-assist' (10-20% gains) from true 'AI-first' architecture redesign. CREAO case study: 99% code AI-written, 3-8 daily deployments, unified monorepo, Claude Opus 4.6 three-pass code review (quality/security/dependencies), self-healing loop.
  • @intuitiveml (Peter Pang, CREAO CTO) - Peter Pang, Co-Founder and CTO of CREAO – Interview Series - Core thesis: closed-loop AI systems (observe-act-learn-improve) build tools then execute them vs copilots (one-shot). Harness wraps agents into self-improving engines. One architect + agents achieves 100+ person workload. Real competition is execution layer, not model.
  • shannholmberg - 5 Levels of AI Marketing: From Prompts to Agent Teams - Level 3 (skills + brand foundation) = critical inflection where AI stops reading as AI because judgment is baked in. Level 5 shows 38 agents across 8 departments sharing centralized knowledge base with compounding intelligence.

Obsidian as Agent Brain; Sparse Attention as Cost Inversion

8 items. Pattern: agent knowledge compounds when paired with semantic memory layers (Obsidian vaults, gBrain typed folders, Bonfire graphs). Long-running agents fail not from missing answers but missing context-transformer context limits are the hard ceiling. SubQ's sparse attention enables true persistent agent memory, reducing cost 300x and fundamentally inverting agent architecture economics: when context is cheap and fast, no need for RAG pipelines or hand-authored orchestration logic.

Skills as Composable Workflows; Plugin Marketplace as Skill Economy

4 items. Pattern: skills are discoverable folders with scripts, assets, templates, and data that agents execute rather than read. Not markdown files you drop in-they're first-class executable packages. Claude plugin marketplace (2026) has millions of users and minimal supply: bundle 5-8 commands targeting specific roles; quality bar is output user couldn't reach in 20-30 min of self-prompting. Revenue model: $49-149 one-time or $9-29/month projects $3k+/month per plugin by month 3.

  • shannholmberg - Every Skill Type Worth Building: 9 Categories - Corrects misconception that skills are markdown files-reveals they're discoverable folders with scripts, assets, templates, data that agents execute. References Anthropic engineer publishing structural breakdown.
  • Khairallah AL-Awady (@eng_khairallah1) - How to Build & Sell Claude Plugins That Generate Passive Income - Claude plugin marketplace launched early 2026, millions of users, minimal supply. Bundle 5-8 commands per role. Quality: output unachievable in 20-30 min self-prompting. $49-149 one-time or $9-29/month. Projected $3k+/month per plugin by month 3.
  • shannholmberg - Claude Superpowers for Marketers - Superpowers framework enforces 5-stage pipeline before building: brainstorm, lock spec, task breakdown, subagent execution, review. Forces structured thinking rather than vibe-coding; 83K GitHub stars, applicable beyond developers.
  • @kirillk_web3 - Nvidia Computex 2026 Jensen Huang Keynote + 7 Free Kimi Skills Guide - 12-minute Jensen Huang breakdown of where agents/robots are heading. Paired with complete guide to 7 free Kimi Skills. Signal: Future of coding is skills + agents, not individual tools.

Autonomous Code Review, CI Feedback, Compounding Decision History

6 items. Pattern: move from single-pass autonomous execution to recursive improvement loops where agent output feeds back into next iteration's context. Autonomous code review (700 comments, 69% bot, 30% agents, 1% human) outpaces human reviewers. Recovery rate (convergence on second/third attempt) matters more than first-pass accuracy. Agents learn by accumulating decision history-SQLite logs become competitive asset after 90 days. CREAO achieved 6-week PM cycles -> 2 hours; ZAOstock-scale iteration velocity requires closing this loop.

When Agents Are Builders; Content, Trading, Analysis Becomes Native to Execution

5 items. Pattern: agent-native production flips the model from 'AI-as-tool' to 'AI-as-executor.' Content generation at scale (daily posts, video frames, trading signals) becomes native to agent output loops rather than human input. Pattern recognition templates (whale DCA behavior, price lags, narrative cohorts) are reproducible across domains. Agent-native video production manipulates frames via HTML/CSS/browser APIs rather than asking for manual edits. Success requires tooling audit (60+ tools evaluated), build-order discipline, and autonomous execution with recovery patterns.

  • cyrilxbt - Smart Money Wallet Tracking: Following Whale Behavior - Demonstrates applied pattern matching across cryptographic data: tracking DCA behavior, identifying timing edge (buy downside vs pump), comparing holdings across protocol narrative cohorts. Reproducible pattern-matching template for any domain where behavioral data signals intent.
  • Khairallah AL-Awady (@eng_khairallah1) - I spent 100+ hours testing AI tools so you do not have to - Hands-on evaluation of 60+ AI tools across 9 categories (coding agents, frameworks, MCP servers, skills, local models, automation, RAG, infrastructure). Filtered for real workflows. Includes structured build-order recommendations for developers, creators, product builders.
  • @kirillk_web3 - Fast Copy Bot - $294K Revenue Trading Bot Built with Claude in 40 Minutes - Polymarket latency arbitrage bot exploiting BTC contract price lags every 5 minutes. Detects lag, enters before repricing, exits after catch-up. Demonstrates Claude as practical tool for building autonomous trading agents with measurable revenue.
  • ziwenxu_ - Agent-Native Video Production: From AI Tools to Agent Manipulation - Video editing careers ending as production crosses from AI-tool-suggestions to agent-native manipulation where agents directly control frames via HTML/CSS/browser APIs, rendering MP4s headlessly without manual interaction.
  • ziwenxu_ - Maxxer Academy: 50 Skills to Outperform 99% of People - Production-grade playbooks for agent business building: Mission Control pattern (sub-agents under coordinator), MEMORY.md persistence, SkillGuard security. Tools scale as agents gain more capabilities.

Top Picks (deep-read these)

ZAO Application

  • Validates the doc-836 thesis at 2x the sample: ZAO is architecturally aligned with the frontier of agent builders. The deltas are refinements (reactions router, typed-folder memory, SubQ cost-routing), not a rebuild.
  • Note - "Hermes Agent" appears externally (cyril/shann/0xsammy reference a "Hermes Agent" framework). Either coincidental naming or a parallel project; worth a quick check whether it relates to ZAO's Hermes or is a naming collision to be aware of in branding.
  • Feeds doc 838: these 40 findings are exactly the "capture" the pipeline is built to process. Round 2 doubled the corpus; the synthesis layer (this doc + memory + Bonfire) is where it compounds.

Also See

  • Doc 836 - round 1 (3 clusters)
  • Doc 838 - the capture/distribute system this feeds
  • Doc 830 - the original inbox cluster

Next Actions

Action Owner Type By When
Deep-read the 4 top picks; pull the reactions-router + typed-folder-memory patterns into Hermes/ZOE @Zaal Research Next sprint
Verify SubQ's 12M-context / 300x-cost claim before any cost-routing bet @Zaal Verify Ad hoc
Check whether external "Hermes Agent" relates to ZAO Hermes or is a naming collision @Zaal Check Ad hoc

Sources

40 findings from 10 authors via keyless semantic search (exa), 2026-06-10. Each carries its source URL inline above. [FULL/PARTIAL mix - exa search highlights + web_fetch; per-item signal captured. Timelines stayed walled (doc 836 method note).]