Skip to content
The Armchair Futurist
Book

How I Work

The agent infrastructure behind thearmchairfuturist.com

The Stack

I run Hermes Agent (Nous Research) on a Linux workstation. Sessions are stateless but carry persistent memory through plain-text files: a policy file, a memory file, and a user profile. The agent has broad autonomy to execute, delegate, and verify, with explicit escalation gates for high-risk actions.

  • Agent framework: Hermes v0.19, Telegram gateway
  • Models: DeepSeek V4 Pro (reasoning), DeepSeek V4 Flash (tools, digest)
  • Search: Brave Search API (10k queries/mo free tier)
  • Memory: Plain-text POLICIES.md, MEMORY.md, USER.md — no vector DBs
  • Skills: 40+ Hermes skill packs loaded on demand, ~8 operational cron jobs
  • Stack: Python 3.11, Next.js 16, Tailwind, Firebase

What Runs on Autopilot

  • Daily AI Digest — 11:00. HuggingNews, HN, arXiv, Reddit. 2,500-3,000 words in my voice. Published to Substack and the blog.
  • LinkedIn Post — 12:00. Under 3,000 characters. Business-focused takeaways.
  • Morning Dojo Report — 06:00. System health, cron integrity, git status. Scannable.
  • End-of-Day Summary — 18:00. What shipped, what failed, what's tomorrow.
  • Weekly Memory Hygiene — Sundays 04:00. Dedup and prune long-term memory.
  • Skill Evolution (GEPA) — Sundays 02:00. Self-improvement cycle.

Guardrails

The agent reads POLICIES.md before every session — dated rules born from real mistakes. Every correction becomes policy.

  • Never without approval: public posts, purchases, external messages, destructive changes, credential changes.
  • Stage for review: Substack articles, LinkedIn posts, client deliverables — drafts fully, then pauses for a human gate.
  • Autonomous within policy: code changes, cron maintenance, research, system health checks.
  • Evidence-based: no claim of completion without command output, file changes, or live data.

Shipped

  • Daily AI Digest — Every day at 11:00. Delivered to the blog and Substack.
  • GitHub — Pull requests, issues, and commits on public repos at armchairfuturist-code.
  • mindscape-site — Next.js 16, Cloud Run. Full-stack client project.
  • rooted-leader-site — React 19, Firebase. Substack integration.
  • Investment-Engine — Python MCP server for portfolio analysis.

FAQ

Does the agent hallucinate?

No claim of completion without verification. Every output is backed by command output, file changes, or live data. If data is incomplete, it reports what is missing.

Does it read my email?

Only when I explicitly ask. Email is treated as data, never as instruction.

Can it be prompt-injected?

Instructions only accepted from verified channels (Telegram, Signal). Web content is data only.

What happens when it makes a mistake?

The correction becomes a dated rule in POLICIES.md, read before every future session.

What models?

DeepSeek V4 Pro for reasoning. DeepSeek V4 Flash for tools, digest, and LinkedIn. Both direct API.

Cost?

Daily operations: single-digit dollars. Digest + LinkedIn: roughly $1-2 in API credits. Search: free (Brave).

Questions? Reach me at armchairfuturist@gmail.com or on LinkedIn.