solo.io

Service mesh started by securing microservices. Now, it's becoming the backbone for AI and agentic traffic. Enterprises are seeing new AI communication patterns: agents calling LLMs, MCP servers, and other agents. These interactions challenge many assumptions we made for traditional microservices traffic. Routing, security, and policy decisions now depend heavily on context inside the request itself. Wrote a blog about how Istio and agentgateway are evolving into an agentic mesh, with concrete examples.

Digging into the 'act' claim on token exchange (RFC 8693) and why it (by itself) needs more help: authorization attenuation, cross-boundary, channel bound tokens, and cryptographic chains. Great blog post from WorkOS describing what's missing, what work is going on in OAuth working group, and what can be done now to help: https://lnkd.in/gW8fqDc6

✅ 🤖 🔐 Interested in learning about the #AAuth protocol? I've put together a comprehensive set of end-to-end demos using AAuth, and agentgateway to implement agent identity, user delegation, authorization, and policy enforcement with detailed instructions and working code! 🪪 AAuth is a new protocol written by Dick Hardt with input from many across the industry focused on the challenges of AI agent identity, registration, delegation, user interaction, and access management. Right now it's being worked on in the #IETF through the #OAuth working group. And we are already starting to see implementations of it! AAuth eliminates bearer tokens, implements cryptographic identity, and supports a layered approach to adoption: - identity only, resource/PEP enforced policies - bootstrap that integrates with SPIFFE, WebAuthN, other.. - client/agent registration, across domains (dynamically, etc) - integration with OAuth - person-asserted delegation - mission mandates / boundary checks - cross trust-domain/boundary authorization checks 👇 Take a look at the IETF draft [1], the protocol explorer [2], and the full-working demo I've put together [3]. See all links in the comments. 🙏 #aauth #oauth #identity #mcp #modelcontextprotocol

👉 🤯 A common source of confusion when working with customers: understanding AI agent "on behalf of", OAuth, and what things look like in practice. OAuth was built for user-to-client delegation, how does that play out in the world of AI agents? Take a look at my blog post here which dives deeply into the topic: https://lnkd.in/guvpYPZ8

One of the top integrations people keep asking about: “How do we make our existing APIs usable by AI agents without rebuilding integrations from scratch?” With Agentgateway’s new OpenAPI to MCP support, teams can turn existing REST APIs into MCP tools for agents and coding assistants. AI agents can securely use the systems enterprises already have. The hard part of enterprise AI isn’t building agents, it’s securely connecting them to real business systems. https://lnkd.in/eRHBW8Bw

⚒️ 🤖 🔧 If you are connecting APIs and MCP servers to your agents in enterprise environments, 𝐭𝐡𝐨𝐬𝐞 𝐜𝐚𝐥𝐥𝐬 𝐬𝐡𝐨𝐮𝐥𝐝 𝐛𝐞 "𝐫𝐞𝐦𝐨𝐭𝐞" [1]. That is, MCP over stdio with piles of API keys 𝐚𝐫𝐞 𝐚𝐧 𝐚𝐧𝐭𝐢-𝐩𝐚𝐭𝐭𝐞𝐫𝐧 [2]. 🤔 But not all MCP servers today are available as remote. And the ones that are, may not be that good. It makes sense to invest in a tool surface/API that is optimized for LLM/AI models. So what are teams actually doing in production today? ▪️ Wrapping existing APIs with thin MCP facades? ▪️ Translating APIs 1:1 into MCP tools? ▪️ Skipping MCP entirely and calling APIs directly? ▪️ Leaning heavily on “code mode” and local execution? 👆 I have a feeling it's the first two bullets? [1]: https://lnkd.in/gTyScmjc [2]: https://lnkd.in/gNP-PfQZ #modelcontextprotocol #mcp #api #claude #openai

Any autonomous agent worth trusting has to answer four questions. Who is this agent? What is it allowed to do, and for whom? What is it doing right now? What can we prove afterwards? Identity. Policy. Observability. Evidence. This is the model we’re building the Catalyst around. #ZeroTrustAgents #AgenticAI #TMForum #DTWIgnite2026

Running #NemoClaw (or other harnesses) on Kubernetes hasn't been that straight forward. In this video, I walk through our recently contributed support to multi-tenant, secure OpenClaw (and soon to be Hermes) running on #Kubernetes with kagent . Blog with more details in the comments!

🔥 MCP can be a bloat of a protocol when connecting to tools. But the naive way of just put all of the tools/descriptions/schemas into the context isn't the only way. Some teams have built their MCP servers to use "progressive disclosure" which only exposes a couple tools like find_tool and execute_tool, but what if you don't control the backend MCP server? Or you want to support explicit tool listing? You can wrap your MCP servers with agentgateway as a "composite" MCP server, and let agentgateway implement the progressive disclosure. How does that work? Basically, you configure your MCP servers (with no modification to them to support search/execute, etc) and let clients connect to agentgateway. Agentgateway will respond to "tools/list" with only two tools: 'get_tool' and 'invoke_tool'. The model can then decide which tools it wants the full schema for. You can combine this with a simple front matter in your prompt and reduce token bloat from MCP by at least 90% Take a look at Michael Levan recent blog post on the setup and cost savings of this approach with agentgateway https://lnkd.in/g8GbQwah #mcp #context #openai #anthropic #progressivedisclosure