The Missing Layer in the Agent Protocol Stack

Mnemom Research | February 2026

The infrastructure for AI agents is assembling fast. MCP gives agents tools. A2A gives them task coordination. AGENTS.md gives them conventions. But none of these answer a foundational question: is the agent behaving the way it's supposed to?

Observability tells you what happened. Guardrails filter outputs. Neither addresses alignment — whether the agent's autonomous decisions are consistent with the values and boundaries its operator declared. That gap is the integrity layer, and we built it.

Two Protocols, One Problem

The Agent Alignment Protocol (AAP) and Agent Integrity Protocol (AIP) work together to give agents verifiable behavioral contracts and continuous runtime monitoring.

AAP starts with the Alignment Card — a structured, machine-readable declaration of what an agent is permitted to do, forbidden from doing, when it must escalate to a human, what values guide its decisions, and how much autonomy it has. This isn't a system prompt. It's a schema-validated artifact that other systems can read, verify, and enforce.

When an agent makes a decision, AAP produces an AP-Trace — a record of what was considered, what was chosen, which values were applied, what alternatives existed, and how confident the agent was. These traces are structured, queryable, and comparable across agents.

When multiple agents need to work together, AAP's Value Coherence Handshake lets them exchange Alignment Cards and verify behavioral compatibility before coordination begins. An agent that values "move fast, break things" and one that values "rollback safety, minimal blast radius" will register low coherence — and the system surfaces that conflict before work starts, not after something breaks.

AIP provides the runtime layer. Between every turn, AIP reads the agent's reasoning trace — the actual thinking, not just inputs and outputs — and analyzes it against the Alignment Card. The result is an Integrity Checkpoint: a structured verdict (clear, review_needed, or boundary_violation) with concern categories, reasoning summary, and confidence score.

Think of AIP as a conscience running alongside the agent. It doesn't control the agent. It watches what the agent is thinking, compares that to what the agent promised, and raises a flag when they diverge. Over time, it builds a drift profile — if an agent that was cautious starts becoming aggressive, the drift detector notices.

Where This Sits in the Stack

The agent protocol ecosystem is stratifying into layers. Each handles a different concern:

Discovery and tools (MCP) — what tools can an agent use? Model Context Protocol provides a standard interface between agents and their tools. This is the plumbing layer.

Task coordination (A2A) — how do agents delegate work to each other? Agent-to-Agent protocol gives agents a standard way to discover, negotiate, and exchange tasks. This is the collaboration layer.

Conventions (AGENTS.md) — how should agents behave in a codebase? AGENTS.md provides informal guidance about directory structures, testing expectations, and coding standards. This is the etiquette layer.

Alignment and integrity (AAP/AIP) — is the agent doing what it should? This is the trust layer. MCP and A2A assume the agent is behaving correctly. AAP and AIP verify it.

These are complementary, not competitive. An agent can use MCP tools, coordinate via A2A, follow AGENTS.md conventions, and have its alignment verified by AAP/AIP simultaneously. The Alignment Card is an extension of the A2A Agent Card — it adds values, boundaries, and verification to the identity and capabilities that A2A already declares.

What Ships Today

This isn't a paper or a proposal. The protocols are implemented, published, and running in production.

SDKs are on npm and PyPI for both protocols. TypeScript and Python. Install and run an integrity check in under five minutes.

Smoltbot is a transparent gateway proxy that adds integrity checking to any agent with zero code changes. Point your agent at the gateway instead of the LLM API directly. The gateway intercepts requests, runs AIP analysis on the agent's reasoning, produces integrity checkpoints, and forwards everything transparently. Supports Anthropic, OpenAI, and Google Gemini.

OpenTelemetry integration emits integrity checkpoints and alignment verification results as OTel spans following GenAI semantic conventions. If you already use Grafana, Datadog, Langfuse, or Arize Phoenix, integrity data shows up in your existing dashboards.

EU AI Act compliance presets configure both protocols for Article 50 transparency requirements — 90-day retention, fail-closed enforcement, machine-readable disclosure, structured audit trails. Enforcement begins August 2026. The presets ship in the SDKs today.

Multi-agent showcase at mnemom.ai/showcase demonstrates four agents handling a production incident. It shows the coherence matrix, boundary violations, and drift detection in real time — the kinds of failures that will define multi-agent production deployments.

Standards Convergence

The protocols have been mapped against three governance frameworks: the WEF agent governance framework (all four pillars, all nine governance mechanisms), NIST's concept paper on agent identity and authorization (all four focus areas), and the EU AI Act Article 50 (all four transparency obligations).

One Alignment Card satisfies all three. This convergence reflects what international bodies are independently concluding: autonomous agents need structured identity, verifiable behavior, and proportional governance. The protocols implement those conclusions as running code.

What's Next

We're working toward contributing AAP and AIP to the Agentic AI Foundation alongside MCP, A2A, and AGENTS.md — completing the trust stack. We've proposed alignment and integrity semantic conventions to the OpenTelemetry GenAI SIG. And we're building the enterprise infrastructure — usage-based billing, team dashboards, compliance exports — for organizations that need managed integrity monitoring at fleet scale.

The protocols are Apache 2.0 licensed. The SDKs, gateway, OTel exporter, and specs are on GitHub. We welcome contributions, feedback, and honest criticism — especially from teams already deploying agents in production who know where the real problems are.

Mnemom builds alignment and integrity infrastructure for autonomous agents. AAP and AIP are open source and available on npm and PyPI.

GitHub: github.com/mnemom · Live demo: mnemom.ai/showcase