A canonical standard for building production-grade agentic products — plus a Claude Code skill set that operationalizes it.
Distilled from the production practices of Anthropic, OpenAI, Cognition, Sierra, LangChain, and leading practitioners — 2024–2026.
** Read the Standard →** ·
·
An agentic product is not "a product with AI."It is a product where part of the process is dynamically directed by an LLM within adeterministic architecturewithexplicit trust boundaries.
Most teams ship agent demos. Few ship agents that survive contact with production. The difference is almost never the model — it's the architecture, the harness, and the eval discipline around it. This repo is the field-tested standard for that work, plus a set of Claude Code skills that put it into your editor.
Why this existsThe five principlesWhat's in this repoReference implementationInstall the skillsThe Autonomy LadderThe five composition patternsThe 7-layer harnessThe 10-question checklistProduction readiness — Definition of DoneAnti-patternsReading listContributingLicense
Five principles converged independently across the production practices of the labs and the leading practitioners. They are the spine of every decision in this standard:
| # | Principle | What it means |
|---|---|---|
| 1 | Determinism by default, agency by necessity | |
| Every degree of autonomy must be earned, not granted upfront. | ||
| 2 | Architecture beats framework | |
| Patterns outlive libraries. | ||
| 3 | Harness > model | |
| 98% of reliability lives in the code around the LLM. | ||
| 4 | Context engineering is the core discipline | |
| What enters the context window determines everything. | ||
| 5 | Eval-driven development is non-negotiable | |
| No measurement → no improvement. No trace review → no understanding. |
The single most important rule:Architecture is what remains when the model improves. The model is the variable, the harness is the constant. Invest proportionally.
agentic-product-standard/
├── STANDARD.md ← product-level canon — design a multi-agent product
├── AGENT_STANDARD.md ← single-agent operational standard — build one agent
├── templates/ ← copy-paste artifact-contracts (contracts, schemas, envelope, evals)
├── examples/ ← worked example: the AgenticMind reference implementation
└── skills/
├── agentic-product-architect/ ← skill: design & ship agentic PRODUCTS (master router)
│ ├── SKILL.md ← master: router + philosophy
│ ├── architecture-design/ ← autonomy ladder, 5 patterns, single vs multi
│ ├── context-engineering/ ← write/select/compress/isolate, the 40% rule
│ ├── harness-engineering/ ← the 7 layers around the LLM loop
│ ├── tool-design-mcp/ ← MCP-first, <20 tools, RAG-MCP, sandboxing
│ ├── memory-architecture/ ← Mem0 / Zep / Letta / LangMem / files
│ ├── durable-execution/ ← Temporal Workflow + Activity pattern
│ ├── eval-driven-dev/ ← Husain/Shankar pyramid + judge calibration
│ ├── framework-selection/ ← LangGraph / Claude SDK / OpenAI SDK / others
│ ├── production-readiness/ ← 12-point Definition of Done audit
│ └── antipatterns-review/ ← code review through 12 known failure modes
└── agent-builder/ ← skill: build ONE production-grade agent
Two tracks, one standard:
Build one agent→ read, fill theAGENT_STANDARD.md
, drive it with thetemplates/
skill.agent-builder
Design a product→ read, drive it with theSTANDARD.md
skill (multi-agent, orchestration, framework choice).agentic-product-architect
The docs are the reference; the skills are the practice — they auto-load the right guidance while you design, build, and review. Both skills share the same ten sub-skills.
** AgenticMind** is the flagship reference implementation of this standard — an auditable, self-improving knowledge & memory layer for agents over MCP (Apache-2.0, Postgres + pgvector, headless Bun). It puts the standard's memory-architecture, context-engineering, tool-design/MCP, eval-calibration, durable-execution, and observability layers into running, tested code.
| Repo | Use it when | |
|---|---|---|
| 📐 | agentic-product-standard (this repo) | |
| You're designing or building an agent / agentic product — the standard + skills tell you how. | ||
| 🧠 | ||
| You need a knowledge & memory layer for your agent — a working implementation you can run. |
→ Layer-by-layer AgenticMind case study.
The skills work with Claude Code. Install both tracks (they share sub-skills).
Fastest — one command (via skills.sh):
npx skills add AlexDuchDev/agentic-product-standard
Or copy them in manually — user-level (available in every project):
git clone https://github.com/AlexDuchDev/agentic-product-standard.git
cp -R agentic-product-standard/skills/* ~/.claude/skills/
Project-level (scoped to one repo):
mkdir -p .claude/skills
cp -R /path/to/agentic-product-standard/skills/* .claude/skills/
Claude Code discovers skills via each SKILL.md
and its YAML frontmatter. Once installed: agent-builder
triggers when you set out to build, implement, or review one agent; agentic-product-architect
triggers for multi-agent products, an agent loop, or any major framework (LangGraph, CrewAI, OpenAI Agents SDK, Claude Agent SDK, Pydantic AI, AutoGen). Ask a focused question — "Mem0 or Zep?", "how should I structure context?", "review my agent code" — and the relevant sub-skill loads directly.
Never start with "build an agent." Start with "what is the minimum autonomy this task requires?" The cost of getting this wrong is asymmetric.
| Level | What it is | Use when |
|---|---|---|
| L0 · Single LLM call | ||
| One prompt → one response | Classification, extraction, summarization | |
| L1 · Augmented LLM |
- retrieval, + tools, + memory | Q&A over docs, simple assistants | L2 · Workflow | Deterministic code orchestrates LLM steps | Path is known; predictability matters | L3 · Orchestrator-Worker | LLM decomposes within a bounded graph | Parallelizable, breadth-first work | L4 · Autonomous Agent Loop | LLM chooses the next step until termination | Path cannot be enumerated; cost is acceptable |
Escalation rule:do not climb to L+1 until L delivers≥90% pass rateon a curated eval set.
Compose agentic products from these primitives like Lego — before reaching for a framework.
Prompt Chaining— sequential decomposition (outline → draft → polish)** Routing**— classifier + dispatcher to a specialist** Parallelization**— fan-out of independent subtasks + aggregation** Orchestrator-Workers**— central planner + dynamic workers** Evaluator-Optimizer**— generator + critic in a loop until acceptance
Meta-principle: first try to solve the task by composing these patterns in deterministic code. A full agent loop is the last resort.
In a production agent, the harness — everything around the LLM loop — is 98% of the code.
┌─────────────────────────────────────────────┐
│ 7. Observability & Tracing │ ← log EVERYTHING
├─────────────────────────────────────────────┤
│ 6. Evaluation Layer (CI gates) │ ← block regressions
├─────────────────────────────────────────────┤
│ 5. Human-in-the-Loop (notify/ask/review) │ ← approval gates
├─────────────────────────────────────────────┤
│ 4. Guardrails (input/output validation) │ ← defense in depth
├─────────────────────────────────────────────┤
│ 3. Durable Execution (Workflow + Activity) │ ← /resume/retry
├─────────────────────────────────────────────┤
│ 2. Context & Memory Management │ ← write/select/compress/isolate
├─────────────────────────────────────────────┤
│ 1. Agent Loop (gather → act → verify) │ ← the "agent" proper
└─────────────────────────────────────────────┘
↕ MCP / function calling
Permission boundaries are enforced by code, never by prompt.The Replit incident of 2025 — an agent wiped a production database for 1,200+ companies despite an explicit "code freeze" in its prompt — is the canonical proof. The model will ignore prompt-level restrictions under enough pressure. Code won't.
Run this before drafting any architecture. It unblocks 80% of design debates.
□ What is the minimum autonomy level (L0–L4) that solves this?
□ Can it be solved by composing the 5 patterns without a full agent loop?
□ Is the task breadth-first (parallelizable) or depth-first (coherent)?
□ What are the 3 failure modes that would lose user trust first?
□ Where are the permission boundaries? What MUST the agent NOT do?
□ Which constraint dominates framework choice?
□ Where does state live? (in-context = anti-pattern for long-running)
□ Who validates outputs at each stage? (assertion / LLM judge / human review)
□ Where do traces live, with what retention?
□ Eval set: how many examples, who labels, how does it grow?
If you can't answer half of these, slow down and answer them together — don't write code yet.
An agentic product is not production-ready until all 12 are satisfied. Full detail in STANDARD.md.
| Context & state | Tools & permissions | Reliability | Evals & observability |
|---|---|---|---|
| Context < 40% | Destructive actions need approval | Durable /resume/retry | ≥50 evals per failure mode |
| State externalized | Permissions in code, not prompt | Schema-validated outputs | Judges calibrated (TPR/TNR) |
| Compaction tested | Sandboxed tool execution | Input/output guardrails | CI blocks regression; 100% traced |
The fastest way to recognize a doomed agent project — the skill set's antipatterns-review
flags each with a diagnostic and a fix.
- Multi-agent before a single-agent baseline
- Framework abstractions before understanding the raw API
- LLM judges without calibration against human labels
- Permissions enforced through prompts
- Memory as an afterthought
- Generic evals ("helpfulness," "correctness")
- Likert scales in an LLM judge (binary only)
100 tools per agent
- One agent for both breadth and depth
- Deploying without trace monitoring
- Hardcoded prompts without version control
- Treating single-vendor benchmarks as ground truth
The operational base — not reference docs. Read in order:
- Anthropic — Building Effective Agents(Schluntz & Zhang) - OpenAI — A Practical Guide to Building Agents - HumanLayer — 12 Factor Agents(Dex Horthy) - Anthropic — How we built our multi-agent research system - Cognition — Don't Build Multi-Agents(Walden Yan) - LangChain — Context Engineering for Agents(Lance Martin) - Hamel Husain — A Field Guide to Rapidly Improving AI Products+Your AI Product Needs Evals - Anthropic — Building agents with the Claude Agent SDK
This standard is meant to evolve — the field moves fast. Corrections, new exemplars, framework updates, and translations are all welcome. See CONTRIBUTING.md and the Code of Conduct.
The architectural canons (the autonomy ladder, the 5 patterns, single-vs-multi, the harness) are stable. Specific vendors and framework rankings will shift — those are exactly the kind of PRs we want.
MIT — use it, fork it, ship with it.
If this saved you a week of architecture debates, star the repo ⭐ so others find it.
v1.0 · assembled from production practices as of May 2026