latestagenticai.com · quarterly tech radar · last reviewed Q4 2026
Latest Agentic AI: The Technology Radar
The live technology radar for the latest in agentic AI (latestagenticai.com): every quarter, each item gets re-argued and moved, added, or retired. Not a leaderboard and not a feed — an opinionated engineering map with reasons, and a changelog of what moved across Adopt, Trial, Assess, and Hold.
Adopt Proven. Use it for real work today.
- MCP for tool connectivity protocols
The de-facto connector standard, now governed by the Agentic AI Foundation, with client support across vendors. Deep dive →
- MCP v2 SDKs, dual-era servers protocols
The 2026-07-28 revision made the protocol stateless. Build new servers on the v2 SDKs (Python mcp 2.x, TypeScript @modelcontextprotocol/server) and serve both eras, because many frameworks still pin mcp<2. Deep dive →
- Agent Skills (SKILL.md) practices
An open format that Claude Code, Codex, Copilot, Cursor and Google's agents all read. Write a procedure once and reuse it across tools. Skills can now also be served over MCP. Deep dive →
- AGENTS.md for repo instructions practices
One vendor-neutral instructions file, maintained under the Agentic AI Foundation and read by most coding agents. Prefer it to a separate file per tool. Deep dive →
- Terminal coding agents tools
Claude Code, Codex CLI and Google's CLIs are ready for daily work. One caveat: from 2026-06-18 Google moved free and personal users from the open-source Gemini CLI to the closed-source Antigravity CLI, so check which one your scripts call. Deep dive →
- Golden-task evals in CI practices
The single highest-leverage practice for agent reliability. Deep dive →
- LLM gateways (routing, budgets, keys) infrastructure
One choke point for cost, access, and fallback across providers. Because it holds every key, it is also the first thing to pin. Deep dive →
- Hash-pinned agent dependencies practices
In March 2026, two LiteLLM releases on PyPI shipped a credential stealer. Pin gateways, SDKs and MCP servers with hashes, and prefer packages that use trusted publishing. Deep dive →
- OWASP Agentic Top 10 as threat baseline practices
A peer-reviewed taxonomy of agent-specific risks (ASI01–ASI10, published December 2025). Map your threat model to it before inventing your own. Deep dive →
- Trajectory logging practices
If you cannot replay what the agent did, you cannot debug or defend it. Deep dive →
- Structured outputs / schema-constrained calls practices
Eliminates a whole class of parse-and-pray failures between steps. Deep dive →
Trial Promising. Use it on projects that can absorb change.
- OTel GenAI semantic conventions infrastructure
Moved down from Adopt. They now live in their own repo, are still marked Development, and renames to the token-usage metrics are queued. Emit them, but pin the version and expect migrations. Deep dive →
- A2A for cross-org agent delegation protocols
The spec is stable since v1.0 (signed Agent Cards, version negotiation) and is now an Agentic AI Foundation project. Real cross-org deployments are still rare, so keep adapters thin. Deep dive →
- AG-UI / agent-native frontends protocols
AG-UI 1.0 packages shipped (Python and TypeScript) and accept 0.x streams. Streaming agent state beats chat-only UX, but governance is still led by one vendor. Deep dive →
- MCP Apps (interactive UI from tools) protocols
Moved up from Assess. It is now an official extension: tools return sandboxed HTML views as ui:// resources. Host support varies, so always ship a text fallback.
- CIMD for MCP client registration protocols
The 2026-07-28 spec deprecates Dynamic Client Registration in favour of Client ID Metadata Documents. New remote servers should accept CIMD and keep DCR only for older clients. Deep dive →
- Durable-execution agent frameworks frameworks
Resumable long-running trajectories are the strongest reason to adopt a framework. Deep dive →
- Microsoft Agent Framework frameworks
GA since April 2026 (1.0) and the successor to AutoGen and Semantic Kernel. If you run either of those, this is your migration path. It speaks MCP and A2A. Deep dive →
- Pydantic AI 2.x frameworks
Version 2.0 (June 2026) reorganised the API around capabilities. It is a good typed, model-agnostic option for Python teams. V1 gets security fixes only. Deep dive →
- Local serving for agent workloads models
llama.cpp- and vLLM-class serving is production-grade. llama-server now speaks the OpenAI Responses and Anthropic Messages APIs and can host stdio MCP servers itself (experimental). Model capability is the constraint. Deep dive →
- Langfuse v4 infrastructure
Observations become the queryable unit, and evals can run per observation. It breaks older SDKs, and Langfuse Cloud retires the v3 endpoints on 2026-11-16, so plan the upgrade now. Deep dive →
- Computer-use agents for narrow workflows practices
Works for constrained, verifiable tasks; supervise anything open-ended.
- LLM-as-judge with human calibration practices
Scales fuzzy evaluation — if you routinely audit the judge against humans. Deep dive →
Assess Watch it. Prototype if it matches a real need.
- MCP Tasks extension protocols
In 2026-07-28 Tasks moved out of the core into an official extension that clients must opt into. The Python SDK does not implement it yet, so prototype long-running calls before you rely on them. Deep dive →
- Small models as tool-routers models
Cheap local routing in front of frontier planners; promising cost profile.
- Decision-model endpoints models
llama.cpp and Ollama both added an endpoint for models that return choices with probabilities, aimed at routing and triage. Too new to build on; watch it.
- Agent payment protocols (AP2, ACP, x402) protocols
Several competing protocols and frequent pivots: AP2 went to the FIDO Alliance and ACP refocused on product discovery. Pilot only, and never without a human approval step. Deep dive →
- Agent memory interchange standards protocols
Everyone rebuilds memory; portable standards are early but worth watching. Deep dive →
- Agentic browsers frameworks
Powerful demos; security model against injection still being proven. Deep dive →
Hold Proceed with caution — known traps at current maturity.
- Deprecated MCP features in new code protocols
The 2026-07-28 revision deprecates Sampling, Roots, Logging and the HTTP+SSE transport, and HTTP+SSE can be removed in the next revision. Don't build new features on them. Deep dive →
- Fine-tuning to inject knowledge models
Wrong tool for facts: no provenance, no erasure, slow updates. Retrieve instead. Deep dive →
- Prompt-only security practices
"We told it not to" is not a control. Enforce at the tool layer. Deep dive →
- Unsupervised writes to systems of record practices
Irreversible + autonomous + unaudited is how agents make the news. Deep dive →
- One mega-agent with every tool frameworks
Context bloat, contaminated trajectories, unreviewable capability surface. Deep dive →
What changed this quarter
Q4 2026 (reviewed 2026-10-08)
- New in Adopt: MCP v2 SDKs with dual-era servers, Agent Skills (SKILL.md), AGENTS.md, terminal coding agents, and hash-pinned agent dependencies.
- OTel GenAI semantic conventions moved Adopt → Trial: they now live in their own repo, are still marked Development, and token-usage metric renames are queued.
- MCP "beyond tools" retired and split: MCP Apps → Trial, the Tasks extension → Assess, and deprecated features (Sampling, Roots, Logging, HTTP+SSE) → Hold.
- A2A stays in Trial with a firmer note: v1.0 is stable and the project now sits under the Agentic AI Foundation.
- AG-UI stays in Trial: 1.0 packages shipped, but governance is still single-vendor.
- New in Trial: CIMD for MCP client registration, Microsoft Agent Framework, Pydantic AI 2.x, and Langfuse v4.
- New in Assess: agent payment protocols and decision-model endpoints.
- Local serving stays in Trial: llama-server can now host stdio MCP servers itself (experimental).
Q3 2026
- First edition: 21 argued calls across adopt / trial / assess / hold.
- OWASP Agentic Top 10 adopted as the recommended threat baseline.
- OTel GenAI conventions placed in Adopt with a schema-pinning caveat.
- MCP core in Adopt; its fast-growing edges (apps, UI, elicitation) held in Assess.
newsletter
One practical agentic-AI guide in your inbox. No news, no hype.
Tutorials and decision frameworks as they ship. Unsubscribe anytime.