17 emerging technologies, each on a ring: use now, pilot, evaluate or watch.
Every entry says what it is, why it matters, when to use it and when not to, with its sources. Tap a bead to open its card.
Read on 2026-11-11 from papers, project pages and practitioner write-ups. Vendor figures are treated as claims, not results. Spot something out of date? Send a fix.
1. The 2026-07-28 revision removed the session, so MCP servers scale like plain HTTP services.
2. Agent pilots are common and production is rare; system-level evals are how teams close the gap.
3. System prompts dominate input tokens, yet most calls read nothing from cache.
4. Agents load SKILL.md packages and run code in disposable sandboxes; supply-chain review follows.
Rings
Quarters
1Models & architectures · Pilot
Tabular foundation models.
Predictions on a new table with no training run.
Use it for Churn, pricing, risk and demand on tables up to tens of thousands of rows; a strong baseline before you tune XGBoost.
Models pre-trained on millions of synthetic tables that predict on a new table in one forward pass, by in-context learning: no training, no tuning loop.
Why it matters
Most enterprise data is tables. For small and medium tables they now match or beat tuned gradient-boosted trees out of the box, and some give calibrated uncertainty.
Not for
Very large tables, hard latency budgets, or commercial use under a non-commercial licence: check each model's licence, they differ.
Models that learn the world by predicting meaning, not pixels.
Use it for Robotics, video understanding and planning research; following where non-LLM foundation models are going.
Joint-embedding predictive architectures learn by predicting the representation of missing or future input, not its pixels or tokens. V-JEPA 2 learns from video and plans robot actions from it.
Why it matters
A different route from next-token prediction: models that learn how the physical world behaves from unlabelled video, with very little robot data.
Not for
Text, documents and chat products: there is no production JEPA language stack yet.
Know-how an agent loads only when a task needs it.
Use it for Repeatable procedures: report formats, review checklists, deploy steps. Pair with MCP for live data.
A folder with a SKILL.md file: a name, a description and instructions, plus optional scripts. Agents read only the name and description until a task needs the rest.
Why it matters
Anthropic opened the format on 18 December 2025 and dozens of agents now read it. It is the cheapest way to give an agent your team's procedures without filling its context.
Not for
Installing community skills unreviewed: researchers report security flaws in a large share of them.
One plug for tools, one for agents, one for the UI.
Use it for MCP for every tool you expose; A2A when separate teams own separate agents; AG-UI for live agent UIs.
Three layers that complement each other: MCP connects an agent to tools and data, A2A lets agents delegate to agents, AG-UI streams an agent's work into a user interface.
Why it matters
Build once, plug into many hosts. A2A reached a stable 1.0; AG-UI fills the gap between an agent and the screen.
Not for
Adding A2A for agents that live in one codebase: a function call is simpler.
Use it for New remote MCP servers, and existing ones you need to scale out. Read the migration notes first: the change breaks older clients.
The July 2026 MCP revision drops the protocol session: no initialize handshake and no session id, so each request carries its own version and capabilities.
Why it matters
An MCP server can now sit behind a plain load balancer like any HTTP service, and gateways can route on headers alone.
Not for
Rewriting a working local (stdio) server just to follow the new revision before your clients support it.
Use it for Every agent: budget the window, cache the stable prefix, compact long runs, keep memory outside the prompt.
Deciding what goes into the model's window on each step: retrieved facts, memory, tool results, compacted history, and in what order, so it can be cached.
Why it matters
Quality and cost now depend more on the context than on the model choice.
Not for
Stuffing the whole history or every document into a long window: accuracy drops as inputs grow.
Use it for Every production prompt: fixed order for tools and instructions, variable content last, and a dashboard of cache hit rate.
Building prompts so the stable part (system prompt, tool list, guardrails) is a byte-identical prefix the provider can cache, and watching cache-read tokens in your traces.
Why it matters
Datadog reports that system prompts were 69% of input tokens in its customers' traces in March 2026, and secondary write-ups of its report say only about a quarter of calls on caching-capable models read anything from cache.
Not for
Timestamps, request ids or shuffled tool lists near the top of the prompt: one changed byte and the cache misses.