02The tool map / Emerging tech radar

What's new, and how far to trust it.

17 emerging technologies, each on a ring: use now, pilot, evaluate or watch.

Q3 2026 in brief
  1. 1
  2. 2
  3. 3
  4. 4
Use nowPilotEvaluateWatchModels & architecturesAgents & protocolsServing & contextEvals, ops & safety1234567891011121314151617
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.

2Models & architectures · Watch

JEPA world models.

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.

3Models & architectures · Evaluate

Diffusion language models.

Text written in parallel blocks, not token by token.

Use it for Latency-bound generation: autocomplete, code edits, fast drafts; compare against a small autoregressive model on your own task.

4Models & architectures · Pilot

Hybrid state-space models.

Long context at a fraction of the memory.

Use it for Long-document and high-concurrency workloads on your own hardware; on-device assistants.

5Agents & protocols · Use now

Agent Skills (SKILL.md).

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.

6Agents & protocols · Pilot

The agent protocol stack: MCP, A2A, AG-UI.

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.

7Agents & protocols · Use now

Agent sandboxes.

A throwaway machine for code the model wrote.

Use it for Code interpreters, data analysis agents, coding agents, anything that runs model-written code.

8Agents & protocols · Evaluate

Computer-use and browser agents.

Agents that click and type like a person.

Use it for Internal back-office flows with no API, under human review, in a sandbox.

9Agents & protocols · Pilot

Durable execution for agents.

Long agent runs that survive crashes and pauses.

Use it for Long-running or multi-step agents with side effects, and human approvals in the middle.

10Agents & protocols · Pilot

Stateless MCP (2026-07-28 revision).

MCP servers that scale like any web service.

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.

11Serving & context · Evaluate

Disaggregated prefill and decode.

Separate GPUs for reading the prompt and writing the answer.

Use it for Large self-hosted fleets with long prompts and many users.

12Serving & context · Use now

Context engineering and memory layers.

Choosing what the model sees on every step.

Use it for Every agent: budget the window, cache the stable prefix, compact long runs, keep memory outside the prompt.

13Serving & context · Use now

Prompt-caching discipline.

Keep the prompt's top identical and pay a tenth.

Use it for Every production prompt: fixed order for tools and instructions, variable content last, and a dashboard of cache hit rate.

14Evals, ops & safety · Use now

System-level agent evals.

Score the whole run, not one prompt.

Use it for Any agent going to production: a golden set of tasks, trace-based checks on tool use, a judge model with human spot checks, run in CI.

15Evals, ops & safety · Pilot

OpenTelemetry GenAI conventions.

Traces of model and agent calls any tool can read.

Use it for New tracing work: emit the GenAI attributes now. Keep prompt and output capture opt-in, since they hold sensitive data.

16Evals, ops & safety · Pilot

Prompt optimization against evals.

Let an optimizer tune the prompt against your evals.

Use it for Any prompt that has a golden set and a clear metric.

17Evals, ops & safety · Watch

Agent supply-chain security.

Treat skills and MCP servers as third-party code.

Use it for An allow-list of reviewed skills and MCP servers; pin versions; run untrusted ones in a sandbox.