AgentSmith — Universal Model-Agnostic Agent Operating Harness

Project: PromptPartner/agentsmith (210⭐, Shell) Creator: Lukas Hertig, founder of PromptPartner (former €1.5B exit B2B SaaS founder) License: MIT

Core Thesis

Agent = Model + Harness — the model is ~10% of the outcome; the harness (rules, tools, memory, guardrails, feedback loops) is the ~90%, and it’s the part you control.

This isn’t theoretical — it grew over 6 months of real autonomous work across software dev, devops, marketing outreach, document creation, data crunching, research, design, and general admin, through hundreds of sessions and five iterations.

Architecture

Static vs Dynamic Context

• Static (loaded every turn, paid every turn): a deliberately lean CLAUDE.md assembled from the core (universal rules) + exactly the profile(s) for the current work type
• Dynamic (loaded on demand): other profiles, skills, templates, docs, research, memory — pulled in only when needed
• “Keep the surface small” (Rule 10) is load-bearing: every line added to static rules makes the agent a little worse at everything else

9 Work-Type Profiles

software-dev · devops-setup · marketing-outreach · document-creation · data-crunching · general-admin · deep-research · creative-design · autonomous-loops

Each defines: what done and verified mean, quality gates, failure modes, STOP-table addendum.

10 Principle Rules (Universal — never relaxed by profiles)

1. Understand before you change (Chesterton’s Fence)
2. Prove it — evidence before assertion (failed-before/passes-after)
3. Verify the whole chain, not just your layer
4. Atomic changes — one concern per commit
5. Verify before you call it done — full checks, not just the one thing you touched
6. Finish the whole change, including the docs (“later doesn’t happen”)
7. Never let a defect evaporate — write it down
8. No live secrets in tracked files. Ever.
9. Research and source material is never silently deleted
10. Keep the surface small — every line earns its place

Safety Model

Two Operating Modes

• Conductor — hands-on, step-by-step. Best for debugging, unfamiliar ground, high-stakes edits
• Orchestrator — define a goal, delegate to sub-agents, review outcomes. Best for well-specified features, migrations, parallelizable sweeps

Move between them by task.

Handoff Protocol

At ~25-30% context used (or when user says “handoff”), the agent safe-states and writes a recall prompt so the next session resumes clean.

System-Evolution Mindset

“Most agent failures are configuration failures.”

When the agent stumbles, don’t just fix the symptom — improve the harness so that class of mistake is less likely next time: sharpen a rule, add a gate, a hook, a skill, a feedback note.

Key Evidence Quotes

“The model (Opus, Sonnet, GPT, Gemini…) is the engine. But an engine isn’t a car. The harness — the instructions, tools, memory, guardrails, hooks, sub-agent orchestration, and feedback loops wrapped around the model — is the car, the road, and the traffic laws.”

“Most agent failures are configuration failures. When a session goes sideways, the honest cause is rarely ‘the model is dumb’ — it’s a missing rule, a vague instruction, an absent guardrail, a tool that wasn’t reached for, or a context window full of noise.”

“A coding agent moved from outside the Top 30 to the Top 5 on Terminal Bench 2.0 by changing only the harness.”

Relevance to Hermes Ecosystem

• Harness philosophy directly validates Hermes Agent’s architecture direction — Manas as orchestrator, Alaya/Anchor/Mano as specialized agents, all governed by shared rules
• Prove-it rule (evidence before assertion) should become a Hermes hard rule — no “I think it works” without actual output verification
• Static vs dynamic context informs how we structure Hermes profiles (CLAUDE.md = lean core, skills/plugins = dynamic)
• Conductor vs orchestrator maps to Manas (orchestrator) vs Mano (conductor for browser tasks)
• Safety model (cautious/trusted) is directly applicable to Hermes pre-exec guard and permissions
• Handoff protocol at 25-30% context is actionable for Hermes context compression decisions
• System-evolution mindset aligns with the skill improvement loop in knowledge digestion

Actionable Recommendations

1. P0: Adopt the “Prove It” rule in Manas’s operating model — never claim success without evidence (test output, screenshot, verified result)
2. P1: Evaluate AgentSmith’s safety model for Hermes permission system — cautious/trusted mode per profile
3. P1: Apply static-vs-dynamic context design to Hermes profile optimization — what goes in CLAUDE.md vs what’s loaded on demand
4. P2: Port the handoff protocol to Hermes context compression — proactive save-state at context thresholds
5. P2: Integrate the Stop-Table anti-rationalization pattern into Hermes’s verification-before-completion skill