Frontier agent ecosystem • comparison

Gemini ADK, MCP/Skills, ChatGPT Agent Mode — compared through the CRSP Type paradigm

Most “agentic” systems optimize execution. CRSP’s Type paradigm distinguishes something else: whether an agent can justify its right to act and release assertions under governance. This page maps frontier stacks into Type-I (non-normative), Type-II (normative), and Type-III (institutional normative).

Type-I

Non-normative executor: produces outputs and performs tasks without entitlement logic.

Speed & convenience Tool execution No entitlement
  • Can: generate drafts, call tools, complete workflows
  • Cannot: justify why an assertion is institutionally permissible
  • Risk: “well-logged wrongness” (trace without defensibility)

Type-II

Normative agent (local scope): bounded commitments with entitlement and evaluation before action.

Commitment Entitlement checks Refuse / revise
  • Requires: Commit → Entitle → Evaluate → Release (local)
  • Strength: task-scoped norm compliance and refusal behavior
  • Gap: not automatically audit-survivable institutionally

Type-III

Institutional normative agent: AER gates, defeaters, policy versions, accountable roles, and audit-ready artifacts.

Ask → Evidence → Release Defeaters Policy-versioned trace Human accountability
Ask: normalize intent into bounded artifact request (audience, posture, prohibitions, accountable roles) Evidence: gather only from approved sources; bundle + reference; flag gaps Release: defeater checks; approve-for-review / downgrade / refuse; log policy version + evidence bundle ID
Type-III is not “more autonomy.” It is less autonomy, more answerability: outputs are drafts for accountable humans to approve.
System Native CRSP Type What it optimizes What’s missing to reach Type-II / Type-III
Gemini ADK
Agent/workflow framework		 Type-I Orchestration as code: structured workflows, routing, retries, tool execution patterns. Type-II: explicit commitments + entitlement/evaluation rules.
Type-III: AER gates, defeaters, policy versions, accountable roles, release logs.
Anthropic MCP + Skills
Interop + capability packaging		 Type-I Tool/data connectivity (protocol) and reusable “capability modules” (skills). Type-II: “skills are not entitlements” — add commitment + entitlement checks.
Type-III: evidence bundling + defeaters + governed release to prevent authority leakage.
ChatGPT Agent Mode
Product agent runtime		 Type-I+ End-user task completion: browsing, tool use, operational actions, workspace controls. Type-II: explicit commit/entitle/evaluate/release loop (not just logging).
Type-III: AER governance that produces review-ready artifacts and release decisions.
CRSP / Nested Learning / AER
Normative control plane		 Type-II Type-III Institutional survivability: governance-in-generation, defeaters, evidence linkage, audit-ready trace. N/A — this is the layer that supplies Type-II/III normativity on top of any runtime/framework.

Gemini ADK (mapped)

Strong orchestration substrate. In CRSP terms, it is a high-quality Type-I executor until governed by a normative loop.

Workflow reliability Tooling Add CRSP for release
  • Use it for: deterministic pipelines, agent routing, retries
  • Type-III overlay: route outputs through AER gates before exposure

MCP + Skills (mapped)

Powerful tool surface. In CRSP terms: capability expansion that increases the need for defeaters and evidence discipline.

Interop Tool surface “Skills ≠ permission”
  • Use it for: standard connectors and reusable tool servers
  • Type-III overlay: treat tool outputs as evidence, not authority

ChatGPT Agent Mode (mapped)

Great operator runtime. In CRSP terms: Type-I+ (action-capable) without institutional entitlement and governed release.

Operator Tool execution Needs AER
  • Use it for: end-user workflows, rapid operational assistance
  • Type-III overlay: enforce draft-only posture + release approvals

CRSP + Nested Learning + AER

The governance plane that turns “agents” into accountable institutional speakers: commitments, entitlements, defeaters, and auditable release.

Type-II normativity Type-III survivability Model-agnostic
  • Use it for: regulated drafting, compliance narratives, oversight artifacts
  • Deploy as: wrapper/overlay across multiple runtimes and vendors

Recommended architecture: “stack, don’t choose”

A practical enterprise pattern is to compose a strong executor with a strong interoperability layer, then place CRSP AER gates at the boundary where outputs become institutional speech.

[Runtime / Product Agent] ChatGPT Agent Mode or equivalent | [Orchestration Layer] ADK-style workflows (parallel/loop/routing) | [Tool Interop Layer] MCP servers / enterprise connectors | [CRSP Governance Plane] Commit → Entitle → Evaluate → Release Ask → Evidence → Release + defeaters + audit log | [Institutional Release] approve-for-review / downgrade / refuse (human accountable)
Design rule: treat “tool access” as capability, and “output release” as governance. The normative boundary is the release boundary.