# Evaluate Agent Native

> Evaluate catalog candidates against either the standard five agent-native criteria or the narrow operator-surface track, and check URL Onboarding. Use when asked whether a service, harness, HUD, status line, or control surface belongs in the catalog.

- Skill: `haoruilee/evaluate-agent-native` (Agent Skill)
- Install (CLI): `npx skillmds@latest add haoruilee/evaluate-agent-native`
- Raw SKILL.md: https://api.skillmd.com/api/skills/haoruilee/evaluate-agent-native/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: CC0-1.0
- Author: haoruilee (https://skillmd.com/u/haoruilee)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/haoruilee/evaluate-agent-native

---


# Skill: evaluate-agent-native

Use this skill to select the correct admission track, collect primary-source evidence, and classify a candidate. Most services use the five-criterion standard. Purpose-built surfaces for operating live agents may use the narrow operator-surface track.

## The gold standard: URL Onboarding

Before applying the five criteria, ask the highest-level question:

> **Can an agent join and start using this service by reading a single URL?**

Services that answer YES are exhibiting the strongest possible form of agent-nativeness. They have internalized the agent as first-class user so deeply that the onboarding flow itself is machine-readable:

```
# The full agent onboarding in one instruction:
Read <url> and follow the instructions.
```

Examples:
- **Moltbook**: `Read https://www.moltbook.com/skill.md` — complete registration, heartbeat, posting, DM protocol
- **Ensue**: `Read https://ensue.dev/docs and call POST https://api.ensue-network.ai/auth/agent-register` — shared-memory agent registration
- **autoresearch@home**: `Read https://raw.githubusercontent.com/mutable-state-inc/autoresearch-at-home/master/collab.md` — complete swarm joining, claiming, publishing protocol
- **db9 / mem9 / mails.dev / MailboxKit / Agents Mail**: domain-hosted `skill.md` URL onboarding for database, memory, and email workflows

This is qualitatively different from:
- An SDK that a human developer installs (requires human coding time)
- An MCP server that a human adds to a config file (requires human config edit)
- A REST API that requires API key setup (requires human account creation)

URL Onboarding means the **agent itself** handles all of this — reading, understanding, and executing the join sequence autonomously.

Mark URL Onboarding as a **strong bonus signal** and highlight it prominently in the evaluation report.

---

## When to activate

Activate when the user asks:
- "Is [service] agent-native?"
- "Does [service] qualify for the awesome list?"
- "I want to add [service] — does it meet the criteria?"
- "What's the difference between agent-native and agent-adapted?"
- "Why isn't [service] on the list?"
- "Does [service] have URL Onboarding?"

---

## Choose the admission track

Use the **standard track** for infrastructure an agent consumes or invokes. Use the **operator-surface track** only when the product's original and primary purpose is operating live AI agents.

Do not use the exception for a generic dashboard, IDE skin, terminal theme, or process monitor that merely adds agent labels.

## Standard track: five hard criteria

A standard-track service must pass **all five**. Evaluate each one explicitly.

### Criterion 1 — Agent-First Positioning

**Test:** Does the official homepage or documentation explicitly identify AI agents as the **primary consumer**?

**Evidence to look for:**
- Homepage headline naming AI agents
- Documentation framing agents as the core user
- Product name or tagline that only makes sense for agents

**Red flags:**
- "Now with AI agent support" (agents are an add-on)
- "Build apps, workflows, and agents" (agents are one of many outputs)

### Criterion 2 — Agent-Specific Primitives

**Test:** Does the API expose at least one primitive with **no meaningful human-facing equivalent**?

**Questions to ask:**
- What is the core API object? Agent inbox? KYA token? Claim? Heartbeat? Or generic inbox/token/task?
- Would this primitive exist if agents didn't exist?
- Is the output format optimized for LLM consumption or human reading?

**Pass examples:** agent inbox, KYA identity token, approval gate with context-window injection, `claim_experiment()`, heartbeat protocol, `publish_hypothesis()`.

**Fail examples:** a REST API that sends emails (humans use it too), a webhook any server can receive.

### Criterion 3 — Autonomy-Compatible Control Plane

**Test:** Can an agent complete a full task loop **without a human clicking anything**?

**Questions to ask:**
- Can the agent provision its own credentials?
- Can the agent initiate, execute, and complete the action without a human redirect?
- Does the service provide agent-appropriate constraint mechanisms?

### Criterion 4 — Machine-to-Machine Integration Surface

**Test:** Is the **primary interface** an SDK, REST API, MCP server, webhook, or **machine-readable URL**?

**Questions to ask:**
- Can an agent use this service without a human ever opening a browser?
- Is there a URL, SDK, REST API, or MCP server documented as the primary integration path?

**Note:** A service that exposes a machine-readable `skill.md` or protocol URL (URL Onboarding) passes this criterion with exceptional strength.

### Criterion 5 — Agent Identity / Delegation Semantics

**Test:** Does the service distinguish (a) agent's own identity, (b) delegated user permissions, (c) audit trail?

---

## Operator-surface track

An operator surface must pass **all five track requirements**:

1. **Agent-operations-first:** Official positioning is explicitly about running or supervising AI agents.
2. **Agent-specific live state:** It continuously interprets state such as rollout/session identity, context pressure, tools, subagents, permissions, Skills, MCP configuration, or quota windows.
3. **Session attribution:** Displayed state maps to a concrete agent session, rollout, or typed agent path.
4. **Dedicated operational surface:** It exposes a CLI, wrapper, daemon, hook, or status-line protocol purpose-built for agent operations.
5. **Honest boundary:** It states when autonomy, machine APIs, delegated credentials, approval enforcement, Skills, or MCP are absent. Read-only observation must not be presented as orchestration or authorization.

Session attach/list/stop controls strengthen the case but are not mandatory. An operator surface can be human-facing and read-only; that is the point of this narrow track.

---

## Bonus signals (check all that apply)

| Signal | Weight | Evidence to look for |
|---|---|---|
| **URL Onboarding** ⭐⭐⭐ | **Highest** | Service hosts a machine-readable `skill.md` / protocol doc an agent reads and follows to self-register |
| Dedicated agent identity model | High | Agent gets its own credential/wallet/token |
| MCP server published | Medium | Official MCP server with documented tools |
| Agent Skills (SKILL.md) published | Medium | `npx skills add org/repo` works |
| Per-agent state / memory / session | Medium | State isolated by agent instance |
| Audit / trajectory artifacts | Medium | Machine-readable evidence of agent actions |

**How to test for URL Onboarding:**
1. Look for a `skill.md`, `SKILL.md`, `collab.md`, or similar machine-readable protocol file hosted at the service's domain or GitHub.
2. Ask: could an agent read that URL and complete the full registration/onboarding sequence autonomously?
3. Try the instruction: `Read <url> and follow the instructions` — does it work?

---

## Classification decision tree

```
Is this infrastructure agents consume or invoke?
├── YES → apply all five standard criteria
│   └── PASS → agent-native (standard) ✅
└── NO → was it purpose-built to operate live AI agents?
    ├── YES → apply all five operator-surface requirements
    │   └── PASS → agent-native (operator surface) ✅
    └── NO → agent-adapted, agent-builder, or out of scope

For either qualifying track, add ⭐ when URL Onboarding is real.
```

---

## Evaluation output format

```
## Evaluation: {Service Name}
**Website:** {url}
**Admission track:** Standard / Operator surface

### URL Onboarding Check ⭐
**Has URL Onboarding:** YES / NO
**Onboarding instruction (if YES):** Read {url} and follow the instructions to {join/register/participate}
**Notes:** {what the agent gets by reading that URL}

---

### Criterion 1 — Agent-First Positioning
**Result:** PASS / FAIL / PARTIAL
**Evidence:** "{exact quote}" — {source URL}

### Criterion 2 — Agent-Specific Primitives
**Result:** PASS / FAIL / PARTIAL
**Evidence:** {primitive name and description}
**No human equivalent because:** {explanation}

### Criterion 3 — Autonomy-Compatible Control Plane
**Result:** PASS / FAIL / PARTIAL
**Evidence:** {how agents operate without human confirmation}

### Criterion 4 — Machine-to-Machine Integration Surface
**Result:** PASS / FAIL / PARTIAL
**Evidence:** {URL, SDK, API, MCP details}

### Criterion 5 — Agent Identity / Delegation Semantics
**Result:** PASS / FAIL / PARTIAL / N/A
**Evidence:** {identity model details}

### Operator-Surface Track (complete instead of standard criteria when selected)
1. Agent-operations-first — PASS / FAIL — {evidence}
2. Agent-specific live state — PASS / FAIL — {evidence}
3. Session attribution — PASS / FAIL — {evidence}
4. Dedicated operational surface — PASS / FAIL — {evidence}
5. Honest boundary — PASS / FAIL — {documented limitations}

---

### Bonus signals
- [ ] URL Onboarding ⭐⭐⭐ — agent joins by reading one URL
- [ ] Dedicated agent identity model
- [ ] MCP server published
- [ ] Agent Skills (SKILL.md) published
- [ ] Per-agent state/memory/session
- [ ] Audit/trajectory/replay artifacts

---

### Overall verdict
**Classification:** agent-native ⭐ / agent-native (standard) / agent-native (operator surface) / agent-adapted / agent-builder / out of scope
**Recommendation:** Add to main list / Add to Excluded section / Do not add
**Confidence:** High / Medium / Low
**Reasoning:** {one paragraph summary}

### Next steps
{If agent-native with URL Onboarding: highlight this in the issue and service file prominently}
{If agent-native without: link to issue template}
{If agent-adapted: explain what would need to change}
```

---

## Common borderline cases

### "The product added an MCP server — does that make it agent-native?"

No. MCP support is a bonus signal, not a criterion. The core question is whether the service was **designed from inception** for agents. A human email provider that adds an MCP server is still `agent-adapted`.

### "The service has URL Onboarding but other criteria are weak."

URL Onboarding is the strongest bonus signal but cannot substitute for the selected admission track. It is an amplifier, not a replacement.

### "The service says 'for AI agents' in marketing."

Check the actual primitives. URL Onboarding is a reliable signal because it requires genuine design effort — you can't fake it with a marketing blog post.

### "It is a human-facing Codex HUD. Is that automatically excluded?"

No. Apply the operator-surface track. A HUD qualifies only when it was purpose-built for live agents, interprets agent-specific runtime state, attributes it to concrete sessions, exposes a dedicated operational surface, and discloses its lack of control or delegated authority. A generic terminal dashboard still fails.