Reasoning & Test-Time Compute
Test-time compute is the practice of spending more inference to get better answers: instead of scaling the model, you scale how long it thinks. Reasoning models such as OpenAI's o-series and DeepSeek-R1 emit long chains of reasoning tokens before the final answer (hidden or summarized for the o-series, fully exposed in R1), trained via reinforcement learning to explore, backtrack and self-correct; related techniques include best-of-n sampling with a verifier or reward model scoring candidates, and Monte Carlo style search over solution branches. The empirical result is a second scaling law: on hard math, code and planning benchmarks, accuracy climbs roughly log-linearly with inference budget. The architect's problem is that cost and latency now vary per request by one to two orders of magnitude, so you need routing (cheap model first, escalate on failure), reasoning-effort caps per use case, and a pricing model that anticipates ten-times token consumption on hard tasks.
Voice & Realtime AI
Realtime voice AI replaces the clunky speech-to-text, LLM, text-to-speech pipeline with native speech-to-speech models that consume and emit audio tokens directly. The pipeline approach loses paralinguistic signal (tone, hesitation, interruption) and stacks latencies into two to four seconds; native models served over bidirectional WebSocket or WebRTC realtime APIs bring voice-to-voice latency down toward 300 to 800 milliseconds, close to human conversational turn-taking, and treat barge-in (the caller interrupting mid-sentence) as a first-class event. The obvious enterprise target is the contact center, where the cost of a contained call typically falls by an order of magnitude, but the architecture is unforgiving: you need streaming function calls into backend systems, guardrails that operate on partial utterances, and recording, consent and disclosure obligations that text chatbots never faced. Pilot on low-risk intents; the tail of edge cases in open-ended voice is long.
Agent-to-Agent Protocols & Interop
Agent-to-agent protocols aim to standardize how autonomous agents from different vendors discover each other, negotiate capabilities and delegate work. Google's A2A protocol (now under the Linux Foundation) is the most visible attempt: agents publish an Agent Card, a JSON descriptor of skills, endpoints and auth requirements, then exchange long-running tasks over JSON-RPC with streamed status updates and artifacts. The relationship to MCP is complementary, not competing: MCP standardizes how one agent reaches its tools and data (vertical integration), while A2A-style protocols standardize how peer agents talk to each other (horizontal integration). For an enterprise architect this is 2008-era web services all over again: the spec is the easy part, while identity propagation across agent boundaries, contract testing, and liability when a delegated task fails are unsolved. Adopt MCP now, treat cross-vendor A2A as assess, and design internal agent meshes so a standard can be swapped in later.
Agent Economies & Autonomous Organizations
Agent economies describe a near future in which software agents transact with other agents: negotiating prices, placing orders and settling payment with no human in the loop per transaction. The plumbing is arriving in pieces: payment-scoped credentials and virtual cards, mandate-style authorization where a human signs a spending policy (merchant categories, per-transaction and cumulative limits) that an agent presents cryptographically, and early protocols from card networks and payment providers for agent-initiated commerce. Gartner's machine customers framing captures the strategic angle: if buying decisions are made by agents comparing structured offers, brand and UX matter less than machine-readable pricing, APIs and reliability. The architect's questions are unglamorous and decisive: how is an agent's authority delegated, bounded and revoked; who absorbs the loss on a hallucinated purchase; and does your commerce stack expose anything an agent can actually consume. Start with low-value, reversible transactions and hard budget ceilings.
World Models & Embodied AI
A world model is a learned simulator: a network trained to predict how an environment evolves in response to actions, so an agent can plan by imagination instead of trial and error in the real world. The current frontier treats large-scale video generation as world modeling: systems like OpenAI's Sora line, Google's Genie and Veo, and NVIDIA's Cosmos learn physics, object permanence and cause and effect implicitly from video, and can be conditioned on actions to roll out counterfactual futures. The link to robotics is direct: vision-language-action models are data-starved because real robot demonstrations are expensive, and learned simulators promise cheap synthetic experience plus a substrate for long-horizon planning that token-by-token language models handle poorly. For most enterprises this is watch-list, not roadmap: the near-term relevance is synthetic data for perception systems and simulation for warehouse or plant automation, where the fidelity gap between simulator and reality remains the operative risk.
The Radar: What to Watch This Quarter
A frontier radar only earns its keep when it is run with discipline: a quarterly cadence, a fixed taxonomy, and an explicit ruling of adopt, trial, assess or hold for every blip, in the style Thoughtworks has maintained for years. The mechanics matter more than the content: each item gets an owner, a falsifiable claim (this cuts contact-center handle time by a measurable amount in a pilot), and a review date, and anything that survives two quarters without evidence moves to hold. The failure mode is treating vendor announcements as signal; the corrective is to weight only what you can test, an API you can call, a benchmark you can rerun, a reference architecture you can break. As of mid-2026 my own radar holds test-time compute routing and MCP at adopt, realtime voice at trial, A2A interop and agent commerce at assess, and world models at hold outside robotics-heavy sectors. One page, defensible positions, revisited quarterly.