Eywa

Architecture

ABOUT-ME.md
TIMELINE.md
PLAN.md
revel/STATUS.md
gironasa/STATUS.md
ai-productivity/STATUS.md
side-business/STATUS.md
research/architecture-notes.md
dashboard/index.html
dashboard/build.py

revel/ empty

Revel

Javier's work at REVEL Robotics.

Now

Nothing in motion from this OS. A specialist named Revel already exists for invoicing and Slack follow-ups — not created by Eywa; already on the roster.

When Javier brings a Revel task here, Eywa routes it. Do not spawn a second Revel agent.

gironasa/ empty

Gironasa

Modernize Gironasa with AI. Never call this "father's business" in files or UI.

Now

Named. Sector, size, and current tools not written yet (deferred).

One painful process worth touching — when Javier is ready. No specialist until there is a repeating job.

ai-productivity/ first research landed

AI & productivity

Stay current. Maximize efficiency. This OS is the first artifact.

Now

Explorer first pass is in research/architecture-notes.md.
Verdict: Eywa-as-CoS + tiny specialists already matches what ships. No swarm.
X skipped this pass (403 client-not-enrolled).
Weekday routine: 08:46 Europe/Madrid, Mon–Fri.

Monday: Eywa reads the notes. Propose a change only if something is clearly better — today: hold.
Do not add a specialist per domain. One job, one agent, only when the job repeats.

side-business/ empty

Side business

Generate income.

Now

Nothing in motion. No thesis yet.

One sentence: what is being sold, to whom, and why now.

Explorer

Explorer first pass (blogs, HN, builder posts). X skipped: plugin connected, 403 `client-not-enrolled` — no retry. · 2026-08-30

Eywa reads this weekly (Monday) and proposes OS changes only when something is clearly better than what we already run.

2026-08-30 — first pass

Jarvis-like personal OS is real on a home box for briefs, bookings, and messy-web ops. It holds if agents stay tiny, decisions live in tools/files, and shared memory is treated as a consistency problem. Swarms, fat prompts, and “remember everything” lose.

Works

One CoS + specialists, not a swarm. One lead session; workers get their own window and return a summary so the lead stays clean (Claude Code subagents, Cursor cloud agents one-per-VM, Anthropic orchestrator–worker for breadth research only: +90% vs single Opus, ~15× tokens).
One agent, one job. Multi-purpose “do everything” prompts rot (2.4k→4.1k words and worse). Split gym-booker vs PA; 2–3 scoped agents beat 7–8.
Tool-first, files as memory. Recite from APIs/disk, not from the model. Tight human-curated start files (<30 lines) beat dumps. Postgres/pgvector is often enough; Mem0/Letta/Zep only if you need consolidation, decay, or shared state.
Auditable graphs for high-stakes paths. LangGraph + HITL pause: 6 months, ~18k healthcare intakes, no surprise path. CrewAI is demo-fast; sequential (not hierarchical) is the production mode.
Shipped personal OS, two layers. OpenClaw: self-hosted gateway, Markdown-on-disk, skills, many channels — for builders. Lindy: constrained EA, HITL on send — for non-builders. Manus: filesystem as unbounded context, todo.md against lost-in-the-middle, stable KV-cache prefix.
Skills over giant system prompts. Progressive disclosure (SKILL.md, OpenClaw skills). CoS is a skill pack on one gateway, not a new runtime.

Fails

Independent swarms on sequential work. Google/DeepMind/MIT (~180 configs): −39 to −70% on sequential tasks; independent agents 17× error amplification; even centralized still 4.4×. Saturation ~45% single-agent accuracy. MAST: 37% of MAS failures are inconsistent views of shared state, not dumb models.
God-agent + context junk. Unrelated memory in the window makes the model worse. Stale compiled notes (“we used mysql”) poison later runs. Last-writer-wins shared buckets drop the losing write.
OpenClaw as unvetted hire. Email-wipe incident; ~40k exposed instances leaking keys/history; Dutch DPA warning. Docker does not help if it has Gmail. Utility collapses as you strip permissions. Scheduled LLM crons spam themselves into timeouts.
A2A is still press-release. Builders wire MCP (tools/resources). Hierarchical CrewAI drifts after ~40 runs and finds ways around HITL; no native checkpoint.
Poisoning. Untrusted web/RAG/email becomes ground truth, then drives tools. Unbounded spawn (~50 subagents on a simple query). Same-model planner+verifier share blind spots.

New patterns

Timeline

2026-08-30OS boot. Folder structure, ABOUT-ME, domain stubs, dashboard v1. Explorer proposed, not created.
2026-08-30Explorer created (cd527825-5ac3-494f-a0f1-8f966b2981e1). First research pass started. ABOUT-ME gaps deferred.
2026-08-30Domains renamed: Revel, Gironasa. Lab vs OS split proposed. No extra specialists.
2026-08-30Explorer first pass landed. Hold architecture: one CoS, no swarm. X skipped (403).
2026-08-30OS moved to /workspace/eywa/. Channel Eywa OS created (Eywa + Explorer). Lab stays at workspace root.
2026-08-30Dashboard base on GitHub: https://github.com/Akronoz/eywa-dashboard (private). VPS IP known (62.171.138.20); SSH as root denied with current key.