Hermes

Context

A personal lab where I ran a team of AI agents like a software org — a secretary, a PM, a senior engineer, a QA, a sysadmin — to find the real ceiling of agent collaboration. Three rebuilds across 2026: OpenClaw 1.0 on Telegram, 2.0 on Discord as an event-driven team, then Hermes.

Problem

The arc was the one everyone hits: first the context problem, then hallucination, then the endless loop of tuning prompts to suppress it. Adding persistent memory (Postgres) stopped the agents forgetting, but it did not make them smarter — my busiest agent still fell apart in long sessions, pulling the wrong memories and hallucinating with a confident, self-consistent rationale. Bigger memory, not better memory.

Constraints

A single model coaxes itself into believing its own hallucinations, so nothing could ship on one model’s say-so. Output had to stay trustworthy across several concurrent projects. And the context-window limit underneath was never something I was going to out-engineer.

Approach

• Route around the context limit, not into it The context window is a frontier-lab problem, not something I was going to out-engineer. I rebuilt on Nous Research's Hermes agent framework — whose memory sharpens with use — and designed around the limit instead of grinding on it.
• Bounded scope, cross-model verification Instead of tuning one model's prompts, I gave each agent a narrow scope on a different-spec model and had the models cross-verify each other — the anti-hallucination move.
• Don't over-build against a fast-moving tool I started my own mission board for more data autonomy; halfway in, Hermes shipped its own — board and settings included — so I dropped mine rather than duplicate it. Read the current ceiling first, then spend your energy on the application that fits it.

Outcome

Agents now handle most of the coding, an Opus-builds / Haiku-QA loop does the verification, and my own energy goes to architecture and steering. I stay honest about the ceiling — at scale, agent autonomy is not reliable yet and hallucination is unpredictable. The multi-agent shape I was running by February 2026 turned out close to what Sakana AI later published as Fugu — theirs a trained orchestrator, mine hand-built and months earlier. One product outlived every rebuild: Ivy, a standalone insurance-consultation agent (Sonnet), carried from OpenClaw 2.0 into Hermes. The through-line: don’t optimize what the next model will fix for free — read where the ceiling is, and build the application that fits it.

Stack

Nous Research Hermes · Claude Opus (build) + Haiku (QA) · Gemini · Discord · Postgres · Redis · Python async · self-hosted