A field guide for MCP gateway, agent gateway, and agent registry
from the team building Jarvis AI.
Explore Jarvis AI alongside our independent research on enterprise agentic AI. We are ASCENDING — an AWS Advanced Consulting Partner that builds Jarvis AI, a governance-first, MCP-native agent platform. This is the public research arm of that team: what we've learned shipping agentic systems for enterprises, written for the people doing the same work.
On our desks
Anthropic's MCP changelog, the NIST AI RMF update, and too much coffee.
Four pillars, one argument.
The enterprise AI conversation in 2026 still circles the same four topics. Each pillar here is a long read, not a landing page. Start where your quarter is burning hottest.
Agentic AI
The operating theory of autonomous agents — where they work, where the hype still outruns the evidence, and which workloads actually cost less than the humans.
Model Context Protocol
MCP went from Anthropic research draft to foundation-backed standard in just over a year. A reader-friendly reference for clients, servers, and gateways.
AI Governance
Policy templates, approval workflows, and the uncomfortable organizational questions — written alongside CISOs who already filed their ISO 42001 paperwork.
Enterprise RAG
Retrieval is still the hardest part of the stack. A pillar on document pipelines, re-rankers, evals, and when agentic RAG earns its seat.
The agent control plane, by topic.
Three procurement-shaped topic hubs underneath the MCP pillar. The catalog plane, the data plane, and the protocol-specific gateway pattern that most enterprise stacks settle on by 2026.
Agent Registry
Source of record for every agent and MCP server in the environment — capability schema, ownership, access policy, versioning. The five evaluation questions for vendors.
Agent Gateway
Where every agent invocation passes through. Auth, RBAC, routing, rate limiting, per-call observability. How it differs from an API gateway, and the five questions for vendors.
MCP Gateway
The agent gateway specialized for the Model Context Protocol. Six production vendors in 2026 — selection criteria, vendor landscape, and the federation question.
What we published this quarter
A reader's guide to evaluating MCP gateways
The evaluation criteria we use when readers ask which gateway to pilot: tool-level authorization, credential brokering, per-tool observability, egress enforcement, and policy-as-code. Drawn from the published documentation of the ~15 MCP gateway vendors tracked in this space.
How to measure AI agent ROI without embarrassing yourself
Productivity-minute arithmetic is how the first wave of agent programs embarrassed themselves. A framework from CFO-side reviewers who now require direct P&L impact.
Moveworks vs Glean, after the ServiceNow acquisition
Moveworks closed into ServiceNow at $2.85B in late 2025. A side-by-side rebuilt from public product documentation, Moveworks' and Glean's own homepages, AWS Marketplace listings, and analyst commentary.
Practitioner-written, openly sponsored.
We are not an independent publication. We are the ASCENDING team that ships Jarvis AI — the same people building the gateway, the governance layer, and the MCP integrations we write about. Writing from inside the problem is the point; pretending otherwise would be dishonest and bad for trust.
Every claim is anchored to a public source we can link to — vendor documentation, standards bodies (ISO, NIST, Linux Foundation), analyst reports (Gartner, Futurum), and peer-reviewed papers.
Every page that discusses Jarvis opens with a disclosure. Every comparison that includes Jarvis marks it clearly. We rank Jarvis honestly in our own tables — where it loses, we say so.
Pricing pages are dated. Comparisons show sources column-by-column. When our reading is directional rather than authoritative, we say so on the page, not in a footnote.
Read most this month
Who writes here
Every piece carries a byline and — where the claim is load-bearing — a separate reviewer. Contributors' LinkedIn profiles are linked from every byline for transparent verification.
Founder and editor of Explore Agentic. Reviews every piece on the site: MCP, governance, and enterprise AI buying cycles.
Covers MCP implementation and security: the OAuth 2.1 spec revisions, gateway auth patterns, and tool-calling threat models.
Covers AWS-native agent infrastructure — Bedrock, AgentCore, S3 Vectors — plus the registry and gateway control planes.
Writes the enterprise RAG pillar: chunking, embedding models, re-rankers, and the evals that gate a release.
Covers natural-language data interfaces, healthcare data standards, and the guardrails that keep data agents honest.
Writes retrieval and judge evaluation: recall@k, faithfulness, LLM-as-a-judge calibration, and CI gating.
Covers what AI programs cost to run after signature: staffing models, TCO, seat math, and deployment case studies.
Writes the Jarvis head-to-head comparisons: pricing models, positioning, and where our own product loses.
Edits the case-study library: outcome framing, the baseline behind every metric, and what a customer will stand behind.
Covers AI governance and procurement: ISO 42001, the NIST AI RMF, EU AI Act dates, and the audit trail buyers demand.
Covers pilot-to-production adoption, data sovereignty, and the platform choices that decide whether a pilot ships.
Covers AI procurement: seat minimums, discount line items, renewal escalators, and the NIST AI RMF in contract terms.
Writes the agentic-AI pillar and the agent ROI playbook. Reviews technical claims across the site before publication.
What this hub answers, in plain English.
The six questions our readers — CIOs, AI leads, and platform architects — ask before they bookmark this site. Each answer links into the deeper pillar where it is sourced.
- 01
What is agentic AI?
- Agentic AI describes systems built around an autonomous loop — observe, reason, act, and replan without waiting for a human click. The distinction matters: a workflow with a language model bolted on is not an agent. Real agents can rewrite their own plan mid-execution when a tool call fails or new evidence surfaces. Gartner counted only about 130 vendors shipping anything that meets that bar in mid-2025, against thousands of self-described agentic vendors.
- 02
What is the Model Context Protocol (MCP)?
- MCP is the open standard agents use to discover and call external tools. Anthropic released the draft on November 25, 2024 with reference servers for Google Drive, Slack, GitHub, Git, Postgres, and Puppeteer. On December 9, 2025 it was donated to the Agentic AI Foundation under the Linux Foundation, whose platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI, and which cited "more than 10,000 published MCP servers" at formation. Servers expose tools, resources, and prompts; clients (the agent runtime) consume them through a uniform JSON-RPC interface. The tools specification defines each tool as a name, description, and an
inputSchemathat "MUST be a valid JSON Schema object" — that schema is what makes one integration reusable across every client. Revisions are dated identifiers rather than semver: the current protocol version is 2026-07-28, preceded by 2025-11-25. The practical payoff: one integration per backend, instead of one per agent framework. - 03
How is agentic AI different from enterprise RAG?
- Vanilla RAG retrieves passages from your corpus and stuffs them into a prompt — a one-shot read-and-reply. Agentic systems can chain multiple retrievals, call other tools between them, and decide when they have enough evidence. Agentic RAG (the hybrid) is now standard for any retrieval workload that requires cross-document reasoning. Vanilla RAG is still the right answer for short factual questions where one retrieval will do the job.
- 04
How does AI governance change when agents are autonomous?
- Static review boards do not catch agents that change behavior between runs. Governance has to move to runtime: tool-level authorization, per-call audit trails, policy-as-code that gates tool execution, and guardian agents that supervise primary agents. ISO/IEC 42001 and the NIST AI Risk Management Framework — released as AI RMF 1.0 on January 26, 2023, organised around the Govern, Map, Measure, and Manage functions, and "intended for voluntary use" — both expect this loop for agentic deployments. The protocols are pushing the same way without being able to enforce it: the MCP tools specification says that "for trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations," and that clients SHOULD log tool usage for audit purposes — a client-side recommendation, which is exactly why the enforcement point has to sit in your infrastructure. The compliance question stops being "what model did you use" and becomes "what did the agent actually do, and who approved it."
- 05
Who edits Explore Agentic, and why does that matter for trust?
- This hub is published by ASCENDING — the AWS Advanced Consulting Partner that builds Jarvis AI, a governance-first, MCP-native agent platform. We disclose this on every page that mentions Jarvis or a competitor, and we mark the Jarvis row in every comparison table. Writing from inside the problem lets us show production patterns instead of secondary-source summaries; the visible disclosure is there so readers can weigh the source.
- 06
How often is the hub updated?
- The four pillars — Agentic AI, MCP, AI Governance, Enterprise RAG — are reviewed each quarter and refreshed against the latest standards drafts, vendor releases, and analyst reports. Glossary entries update on demand whenever the underlying spec or category shifts. Every page carries a visible "Updated [month, year]" stamp; the publish and modified dates also appear in the page's Article structured data so search engines see the freshness directly.
- 07
Should we build our own agents or buy an enterprise platform?
- Build for unique workflows where the agent loop is itself a competitive moat — proprietary planning logic, regulated decision paths, or deeply embedded internal tools. Buy for the long tail: support triage, sales research, IT helpdesk, document workflows, employee Q&A. The economics in 2026 favor buy-then-extend: a governance-first platform like Jarvis AI gives you the gateway, registry, audit trail, and MCP integrations on day one, then you author the prompts and tools that encode your specific work. The DIY path adds 6–12 months of platform engineering before the first agent ships, plus the ongoing burden of keeping pace with model and protocol churn. That churn is not rhetorical: the hyperscalers now ship the lower layers themselves — AWS describes Amazon Bedrock AgentCore Gateway as "a single, secure entry point for agentic traffic" handling "both inbound authentication (verifying agent identity) and outbound authentication (connecting to tools) in a single service," and Microsoft gives every Foundry Agent Service agent "a dedicated Microsoft Entra identity" with automatic version snapshots — while a second protocol, A2A, moved to the Linux Foundation with a technical steering committee spanning AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow. Anything you build has to track all of it. Most enterprises that started DIY in 2024 are now consolidating onto a platform.
- 08
How do you measure ROI on enterprise agentic AI?
- Productivity-minute arithmetic — "each agent saves N minutes per task" — is how the first wave of programs embarrassed themselves in 2024–2025. The framework that survives CFO review pins agent value to direct P&L impact: reduced fully-loaded support cost per ticket, deflected hires in a backfill plan, accelerated revenue from faster sales-research cycles, or recovered margin from automated reconciliation. Salesforce reported $1.7M of new sales pipeline generated by Agentforce in year one as a benchmark of the magnitude that lands in board-deck slides. Our /playbooks/ai-agent-roi piece walks through the model line-by-line, including the metrics CFOs reject and the substitutes that hold up.
- 09
Which industries have deployed agentic AI in production right now?
- Customer support and IT helpdesk are the furthest along — Moveworks (now part of ServiceNow), Salesforce Agentforce, and Zendesk's AI agents are all generally available with reference customers reporting 20–40% deflection of tier-one tickets. Financial services have moved on document-heavy workflows: KYC reviews, loan processing, claims triage, with governance layered to satisfy regulators. Healthcare is slower but real for revenue-cycle management and prior-authorization workflows. The pattern: bounded workflows with rich evidence trails and clear undo paths deploy first; open-ended creative or strategic work stays human-led for now.
The team writing this ships Jarvis AI
This hub is the editorial layer. Jarvis is where the patterns we cover — governance, registry, guardrails — get deployed. If you're scoping a program rather than just reading, the product page is the next step.
ascendingdc.com/jarvis-ai — ASCENDING's enterprise agentic AI platform.