create-ai-agent

A disciplined recipe + Hermes skill for building production AI agents the right way — not as one-off prompt loops, but as a wired "organization" you hire, train, and manage out of markdown.

A skill file is an employee. A resolver table is your org chart. An eval is a performance review. The memory is your company brain. You are not writing a prompt — you are building a workforce.

Distilled from 18 talks at AI Engineer 2026 (the aiDotEngineer channel, July 2026) and the companion building-ai-agents-2026 field guide.

---

Why this exists

Most "agent" projects die the same way: a while true loop with no verification, no checkpoint, and no memory hygiene. This skill encodes what the best talks actually recommend:

• Leverage is in the wiring, not the weights. The 2X and 100X founders run the same model.
• Verify deterministically. Types, tests, simulators, ontology/reasoner — don't let the loop grade its own homework with more LLM calls (error compounds across iterations).
• The Save Button. Checkpoint a known-good state before any risky change; rollback on evidence.
• One strong loop beats a multi-agent committee. Split agents only when slices are truly isolated.
• Context engineering > prompt engineering. Compaction is fine; curate the brain.
• Skillify, never one-off. When the agent does something well, turn it into a reusable skill.

---

What's in the box

create-ai-agent/
├── SKILL.md                      # the recipe: 9-step build + rules + anti-patterns
├── templates/
│   ├── AGENTS.md                 # org chart / resolver table (skill file = employee)
│   ├── skill.md                  # one-capability "employee" (typed inputs + verify + rollback)
│   ├── brain.md                  # memory + hygiene (library + librarian)
│   ├── eval-checklist.md         # deterministic + agent-as-judge performance reviews
│   ├── goal.md                   # ultra-goal, editable while running
│   └── checkpoint.sh             # the "Save Button" (snapshot / restore)
├── scripts/
│   └── scaffold_agent.py         # generate a new agent project from the templates
├── README.md
├── CONTRIBUTING.md
├── LICENSE
└── .github/                      # PR + issue templates

---

Install as a Hermes skill

Clone straight into your Hermes skills directory:

git clone https://github.com/Ratinsharma/create-ai-agent.git \
  "$HOME/AppData/Local/hermes/skills/create-ai-agent"

Then load it when you want to build an agent:

skill_view(name='create-ai-agent')

---

Scaffold a new agent

python scripts/scaffold_agent.py \
  --name "support-triage-agent" \
  --purpose "Triage incoming support threads, route to owners, draft replies." \
  --dir "./agents"

This generates:

agents/support-triage-agent/
├── AGENTS.md          # org chart / resolver
├── goal.md            # ultra-goal
├── skills/            # drop skill.md "employees" here
├── brain/README.md    # memory + hygiene
├── evals/checklist.md # performance reviews
└── .checkpoint/       # Save Button helper

On Windows, call python directly (not uv run python3 — that maps to the Store alias and produces zero-byte files).

---

The 9-step build (summary)

1. Write the mandate — one line, one job. Need three sentences? That's three employees; split them.
2. AGENTS.md — resolver table routes work; sets filing rules + context budget.
3. Split latent vs deterministic — mechanics/state → code; taste/judgment → model.
4. Skill files — one per repeated task: typed inputs, verification step, rollback path.
5. Save Button — checkpoint before risky changes; accept/revert on evidence.
6. Verification — deterministic (types/ontology) + agent-as-judge on live traces.
7. Build the brain — provenance on every fact, contradiction handling, pruning.
8. Start at 2–3x — small, verified, human in the architecture. Expand only after trust.
9. Skillify the process — reuse these templates; adapt, don't restart.

---

Source & attribution

Principles distilled from 18 AI Engineer 2026 conference talks, including:

• The Company Brain — Garry Tan (YC)
• The Great Loops Debate — Livingston / Huntley / Horthy / Kostruba
• ActiveGraph / BabyAGI4 — Yohei Nakajima
• SWE Is Not Writing Code — Benoit Schillings (Google DeepMind)
• Codex Workshop — Jason Liu (OpenAI)
• Self-Improving Agent — Jason Lopatecki · Agent as Judge — Aparna (Arize)
• Ontologies — Frank Coyle · Everything Is a Rollout — Alex Shaw & Ryan Marten
• Claude Long-Horizon — Lance Martin · Killed Multi-Agent — ZS · Arch Half-Life — Dan Farrelly
• Skills & New SDKs — Elvin A · The Harness Becomes a Claw — Sam Bhagwat · The Save Button — Hamza Tahir
• Loop Engineering — Kyle Mistele · The Harness Is Not Enough — Dex Horthy · HTML Is All Agents Need — James Russo

Full per-talk crux lives in the companion field guide building-ai-agents-2026 (not published under this account yet). This repo ships the production recipe + scaffolder distilled from those talks.

---

Contributing

See CONTRIBUTING.md. Keep the recipe honest: verify deterministically, no invented APIs.

License

MIT — see LICENSE. Free to use, fork, and ship.

Author

Ratin Sharma — github.com/Ratinsharma