The agentic stack crossed a line in Amsterdam this week.
Two days — September 17-18, 2026 — at RAI Amsterdam and the conversation shifted decisively past demos and into production. The question on every engineer’s lips was no longer “can a model do X?” but “how do we run agents that are reliable, observable, and secure when they hit real infrastructure?”
The RAI Amsterdam auditorium during Day 1 — the conversation had visibly shifted from demos to production reality.
What the data says
The headline numbers came from the 2026 AAIF Member Pulse Survey (n=186), shown in Mazin Gilbert’s opening keynote on Day 1:
- 81% of respondents have agents in production today (51% at production scale)
- 89% use or evaluate MCP, which the slide called “the most adopted agent standard”
- 60% run multi-agent systems, and 50% across more than one stack
- 78% build on open-source tooling, against 22% naming closed-source tools
Outside the survey, the AGNTCon numbers I photographed came from David Soria Parra’s keynote: more than 1 billion MCP tool calls served on Claude, and 950+ MCP servers in the Claude connector directory. AGENTS.md says on its own site that it’s used by over 60k open-source projects; that figure wasn’t on any slide I saw. goose, which Manik Surtani described as a coding harness that came out of Block, is one of the six AAIF projects, next to MCP, AGENTS.md, agentgateway, A2A and Agent Router.
The three shifts that defined the floor
1. From single-agent demos to multi-agent systems
A2A joined the Agentic AI Foundation as a hosted project, and it was on stage from the first keynote: Mazin Gilbert listed it with MCP and agentic commerce protocols as the “open protocols” layer of the Internet of Agents. Day 1 also had a breakout called “A2A Goes Stable: What Changed, Why, and What’s Next” (Sam Betts of Cisco and Kuba Herczyński of Google), and Day 2 had Willem Berroubache (Orange) on multi-agent orchestration in production. With 60% of survey respondents running multi-agent systems, how agents discover and call each other is now a production question.
2. MCP is now infrastructure, not experimentation
David Soria Parra (Anthropic) presented MCP as the default integration layer, with the adoption numbers above. The July 28 specification changes — stateless core, extensions framework, authorization updates — address the production pain points. His slides listed statelessness (no handshake, no sessions, any instance behind a plain load balancer), multi round-trip requests without held-open streams, cacheable results and a twelve-month deprecation window.
Clare Liguori (AWS) built on the same changes from a different angle. Her keynote, “Reactive Agents: Your Agent Doesn’t Need to Be Always On”, was about agents that hibernate while they wait for input or tool results, using stateless calls, the MCP Tasks extension and the Triggers & Events working group’s webhooks to wake up. Her “Infinity experiment” fitted 80,000 resident agents in 8.4 GB of memory, about 103 KB each, against a Raspberry Pi 5 (8 GB) marker.
The gateway pattern (one control point for model, MCP and agent traffic) came up again and again: agentgateway and Agent Router are both AAIF projects, Solo.io’s keynote mapped agentgateway onto the OSI model with a new “L8 Context” layer, and Lin Sun (Solo.io) gave “Keep Infrastructure Out of Your AI Agents: The Agent Gateway Pattern” on Day 2. It isn’t universal yet, though: in the member survey, 15% of respondents named Agentgateway among the standards they use or evaluate.
3. AGENTS.md is how you ship agents that respect your codebase
The message was consistent: an agent that doesn’t know your repo’s build commands, test conventions, and code style is a liability. AGENTS.md solves this with a single Markdown file at the repo root — no JSON schema, no config dance. In the member survey, 51% of respondents use or evaluate it, which put it fourth among open standards after MCP, Skills and A2A.
Key sessions that mattered
- David Soria Parra (Anthropic), “MCP and the Era of Connectivity”: MCP adoption, the stateless 2026-07-28 revision, and what’s next for messaging, skills over MCP, and authorization
- Clare Liguori (AWS), “Reactive Agents: Your Agent Doesn’t Need to Be Always On”: agents that hibernate and wake on input, tool results or external events, with the 80,000-agents-in-8.4-GB Infinity experiment
- Angie Jones (Agentic AI Foundation): host of both days’ keynotes, and on Day 2’s “MCP Challenges & Opportunities” panel with Sam Morrow (GitHub), Shaun Smith (Hugging Face) and Shub Argha (Arcade)
- Dexter Horthy (HumanLayer), “State of the Software Factory” (Day 2): the “lights-off” software factory, the SlopCodeBench results, and “there’s no dark factory without better verifiers”
- Idit Levine and Keith Babo (Solo.io), “Agents as Actors: Harnessing the Power of Agentic Infrastructure”: kagent, agentgateway, and the agentdesktop launch for governing agents on developer laptops
The real problem: context, not compute
My read of the week: many agent failures in production are context problems, not capability problems. Agents lose track of what they’ve done, call the wrong tools, drift from intent, or silently fail. Sam Morrow (GitHub) argued in “MCP Doesn’t Have a Context Problem” that the fix is context engineering in the harness: progressive tool discovery, skills over MCP and code mode.
These patterns were on the schedule (I didn’t attend all of these sessions):
- Agent memory and harness engineering, in Ignacio Martinez’s (Oracle) “Total Recall: Agent Memory and Harness Engineering”
- Agent Router patterns: one integration in front of many model providers and MCP servers, per the AAIF project description, and Ignasi Barrera’s (Tetrate.io) “Your Agents Need a Router”
- Agent evaluation, in Day 2 sessions such as “Testing Agents and Their Tools: Offline Evaluation, Synthetic Tasks, and A/B Experiments” (Ksenia Bobrova, GitHub) and “Evaluating Agents at Scale: From 50 Examples to a Production Flywheel” (Bauke Brenninkmeijer, Orq.ai)
Where this goes next
The free livestream captured all of it. Recordings are rolling out on the Agentic AI Foundation YouTube channel.
If the Amsterdam energy is any signal, agentic systems aren’t coming — they’re already here, and the conversation has moved to making them work reliably at scale. My detailed notes are in the Day 1 and Day 2 recaps.
This is a media-partner recap. My preview article with the full agenda, speaker lineup, and track breakdown is here.
