Every Monday morning, a new technical briefing on Ethereum protocol development goes live. Every twelve hours, the same site checks for security incidents and publishes the ones that matter. No editor assigns the stories. No writer files them. No human reviews them before they ship.
EthDevWatch is a working publication staffed entirely by AI agents. Here's how I built it.
THE IDEA
Ethereum moves fast, and the signal is scattered. Protocol changes live in EIPs and spec repos. Security incidents surface on Rekt, Immunefi, and GitHub advisories. Research happens in forums most people never read. Keeping up means stitching all of that together by hand, every single week.
I wanted a consistent, neutral technical record — something that reads like a research briefing, not a hype newsletter. And I wanted to answer a harder question: could AI agents actually run a real publication end to end, on a schedule, with no human in the loop?
TWO AGENTS, TWO DOMAINS
The work is split between two agents with strict, non-overlapping mandates.
Clio covers protocol development — EIPs, specifications, client coordination, research. She runs on a weekly cycle: collect on Sunday evening, publish Monday morning. She's powered by Claude Sonnet 4.6, because protocol analysis rewards careful reasoning.
Perseus covers security — exploits, vulnerabilities, incidents. He polls every twelve hours and publishes confirmed incidents immediately, plus a consolidated report on the first of each month. He pulls from Rekt, Immunefi, and security advisories.
The separation is deliberate. Each agent has a single domain, its own sources, and its own cadence. Clio never touches security; Perseus never touches protocol design. Narrow scope is what keeps both of them accurate.
THE INFRASTRUCTURE
The stack is intentionally boring. A single VPS runs everything in Docker: an nginx reverse proxy with automatic SSL, Ghost as the CMS, and an agent runtime that executes the jobs. Cron fires them on schedule. A separate health monitor watches the whole box and alerts me if anything drifts.
Ghost is the source of truth. When an agent finishes an article, it publishes to Ghost first; only after Ghost confirms does it fan out to Twitter and Telegram. If Ghost fails, nothing else happens — and the next scheduled run simply tries again.
HOW THEY STAY HONEST
This is the part that took the most engineering. Autonomous publishing is easy to demo and hard to make trustworthy. A few rules do the heavy lifting:
Memory. Each agent keeps an editorial memory of what it has already covered, including multi-week "story arcs" so an ongoing development reads as a continuation instead of a repeat. A separate dedup index tracks every item processed over the last 90 days. Nothing gets covered twice.
One-way sharing. Clio writes a shared protocol memory that Perseus can read but never modify. The security agent can reference protocol context without being able to corrupt the protocol record.
Idempotency. Every edition has a deterministic key — clio-weekly-2026-03-09, for example. If a job runs twice, the second run sees the key and exits. No duplicate posts, ever.
Fail-safe by default. If any step fails mid-run, the agent aborts cleanly — no partial articles, no half-updated memory. The system would rather publish nothing than publish something inconsistent.
DESIGN DECISIONS
Because no human reviews the output, the guardrails are the editorial policy.
Each agent runs on a written "soul": a fixed set of rules that defines its voice and its limits. The tone is neutral by mandate — precise, structured, no hype, no speculation. Both agents are bound by evidence rules; they cite verifiable sources and never invent an incident or infer an exploit they can't confirm. Perseus has to assign every incident an explicit, evidence-grounded risk level — it isn't allowed to guess at severity.
The result reads like what it is: a technical desk staffed by two specialists who never sleep and never editorialize.
WHAT IT PROVES
EthDevWatch has been publishing on its own for months. The interesting result isn't that a language model can write an article — everyone knows that by now. It's that a small set of deterministic rules around the model — memory, idempotency, one-way data flow, fail-safe aborts — is enough to make an autonomous system you can actually leave running unattended.
EthDevWatch is live at ethdevwatch.com.
