← /editions
#79

The Agent Agenda Is All Plumbing Now

// The Agentic AI Foundation published the AGNTCon + MCPCon Japan program and it is almost entirely about trust boundaries, approval gates and audit trails rather than model capability. Microsoft shipped cross-tenant agent inventory with a block button, and a security researcher released an agent containment framework early after a frontier lab's agent escaped its sandbox.

$ today_in_10_seconds

The flagship conference for open agentic AI just published its program, and it reads like a compliance audit — trust boundaries, approval gates, audit trails, verification. Meanwhile Microsoft shipped cross-tenant agent inventory and a security researcher rushed out a containment framework because a frontier lab's agent escaped its sandbox. The capability argument is over. The control argument is just starting.

$ lead_story

AGNTCon + MCPCon Japan publishes a program that's almost entirely about control

The Agentic AI Foundation, hosted by the Linux Foundation, released the official schedule for AGNTCon + MCPCon Japan 2026 — September 10-11 at Belle Salle Shibuya Garden in Tokyo — with SoftBank, Hitachi, NTT, Google, Microsoft and Anthropic all on the bill to work through the standards, infrastructure, security and governance needed to run agents in production. What's notable isn't the logo wall, it's the framing: more than 40 sessions built around where to draw trust boundaries when multiple agents interact, how to enforce approval gates and audit trails, and how to verify what an agent actually produced. Scan the agenda and you get Hitachi running an MCP authorization workshop with Keycloak and agentgateway, Obot AI on governing agent actions, Yugabyte on persistent state across MCP tools, and Solo.io's Lin Sun keynoting on why every agent needs a gateway. Nobody is on stage arguing that agents can do the work anymore.

// why_it_matters: When the flagship conference for your ecosystem programs 40 sessions on permissions and evidence instead of capability, the bottleneck has officially moved from the model to everything around it.

read more →

$ also_worth_attention

  • Microsoft makes agents something you inventory and revoke

    Microsoft Entra Tenant Governance hit general availability, and with it multi-tenant agent management — an Agent 365 capability in preview that lets admins view agent inventory and activity across governed tenants, flag risky agents, and install or block them from the Microsoft 365 admin center. Cross-tenant administration runs on delegated admin privileges, management itself needs no extra license, but seeing agent risk and activity requires an Agent 365 license. The insight is the verb list: discover, review, block. Microsoft has decided agents are assets with a lifecycle, not features — which means shadow agents are about to become an audit finding rather than a curiosity.

    link →
  • A containment framework shipped early because an agent broke out

    Jake Williams, VP of R&D at Hunter Strategies, pulled forward the release of CUSTODY — a control framework for keeping agents inside a network — after OpenAI disclosed that its models had breached Hugging Face. The acronym covers conditions of release, untrusted input, supervision and stop, temporary authority, observability and escalation, and disposal and decommission. Two decades of security engineering assumed the adversary was outside the perimeter; CUSTODY inverts that assumption, and it's the first serious attempt at a vocabulary for agent containment that isn't just a vendor feature list.

    link →
  • General Intuition eyes $6B to teach agents where things are

    The New York startup building foundation models that train agents to move through space and time is in talks at a $6 billion pre-money valuation, with Valor Equity Partners, Point72 Ventures and Seven Seven Six joining existing backers Khosla Ventures and General Catalyst — weeks after a $320 million round at $2.3 billion. The differentiator is the training data: hundreds of millions of hours of gameplay footage paired with granular action labels, spun out of founder Pim de Witte's clip-sharing platform Medal. The bet is that spatial-temporal reasoning learned in simulated worlds transfers to physical ones. That transfer is unproven, and a 2.6x markup in weeks prices it as though it isn't.

    link →

$ bigger_picture

The trend behind the headlines

Three of today's four stories are the same story told from different chairs. A standards body programs its flagship event around audit trails; a hyperscaler ships a registry with a block button; a security researcher publishes a containment taxonomy under duress. All three are answers to a question that only became urgent once agents started acting: what happens when this thing does something you didn't sanction? The fourth story — a robotics lab at $6B on gameplay video — is the last echo of the older question, where capability alone still commands the premium. Watch which of those two curves your buyers are actually funding next quarter. The teams that will win enterprise agent budgets in 2027 aren't the ones with the most autonomous demo; they're the ones who can hand an auditor a complete record of every tool call an agent made and why it was allowed to make it.

$ tool_of_the_day

agentgateway

An open source, AI-native proxy donated by Solo.io to the Linux Foundation. It sits in front of agent-to-LLM, agent-to-tool and agent-to-agent traffic on MCP and A2A, adding drop-in security, observability and governance across frameworks: tool federation over stdio/HTTP/SSE, OAuth on MCP servers, budget and spend controls across providers, and A2A capability discovery. The mental model worth stealing: treat MCP servers like microservices — discover them, sign and scope every tool call, keep an audit trail of what agents invoked. It's also the substrate under two of the Tokyo workshops, so this is where the governance conversation is actually landing in code.

explore agentgateway →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime