The AI field produces more announcements per week than any enterprise can absorb. Most of them do not change a delivery plan. The items below do, sorted by how directly they touch the work on this site: agents in production, autonomous systems, and migrating enterprise systems.

1. Reasoning became a dial, not a model choice

For two years the trade-off was structural: pick the fast model or the smart one, route your request accordingly. That split is collapsing. The frontier families now expose reasoning depth as a parameter — low, medium, high — so the same model handles a classification call and a multi-step analysis, and you pay for the depth you actually needed.

Mistral Large 4 (“Le Chonk”, open weights published 29 October 2026) is the clearest recent example: 1 trillion parameters with 52 billion active, natively multimodal, and three reasoning modes described by Mistral as a way to fit reasoning to task depth without prompt surgery. It also ships with explicit token-efficiency claims — around 47 % fewer tokens per task than a comparable open competitor on coding-agent work. Source ↗

Why it matters for delivery: your cost model has a new variable. If you budgeted per call, you are now budgeting per call × reasoning depth, and the same eval run can cost 3× depending on a flag. Instrument this before you scale anything.

Practical rule: default everything to the lowest reasoning setting that passes your eval set, and promote individual task types upward only when the metric moves. Most "we need a better model" complaints are "we need a better eval set".

2. Open weights reached the frontier — with conditions

The gap between closed frontier models and open-weight releases is now measured in months and, on coding-agent benchmarks, sometimes in rounding error. The 29 October Mistral release is the strongest European data point; the open Chinese and US labs have been shipping comparable models all year.

But "open weights" is a licensing term, not an architecture. Before you build on one, check three things: the licence (commercial use, thresholds, acceptable-use clauses), the hardware floor (the weights may fit, the KV cache and context you need may not), and the tokenizer and tool-calling format, because your agent harness will need to match it exactly.

Why it matters for delivery: for regulated enterprises — and for anyone in SAP or public-sector work — the ability to run a frontier-class model inside a boundary you control is the difference between a project that gets approved and one that stalls in legal. Sovereign deployment stopped being a compromise this year.

3. Agents became the product, and evaluation became the bottleneck

The 2026 shift is not better chat. It is that models are now trained and sold for tool use: multi-step execution, computer use, long-horizon tasks. Benchmarks moved with them — terminal-based coding agents, OS-level computer control, GAIA-style multi-tool reasoning — and the interesting number is no longer accuracy on a single task but how long a task the model can complete unsupervised.

That creates a new kind of engineering problem. If your agent can run for 40 minutes, your failure modes are no longer "wrong answer" but "plausible wrong trajectory", and reviewing the output is no longer enough — you need to review the path. Trajectory monitoring, replayable traces and step-level evaluation are now core infrastructure, not nice-to-haves. Research on real-time intervention in agent trajectories is moving in exactly that direction.

This is why the agent playbooks on this site spend more space on tool contracts, gates and evals than on prompting.

4. Small models got good enough to run where the data is

The most practically useful shift of the year, and the least covered: 3–8 B parameter models, quantised and distilled, now do real reasoning on consumer hardware and single-board computers — sub-second latency, no network, no data leaving the device.

That is precisely what made the reconnaissance drone build practical. A detector that needs a data centre cannot fly. A 6 MB int8 model at 12 fps on a Raspberry Pi 5 can, and the entire privacy argument changes when the pixels never leave the aircraft.

The same logic is reshaping enterprise architecture: classification, extraction and routing tasks that were being sent to a frontier API are moving to a local model on the same machine as the data, with the frontier model kept for the 10 % of cases that genuinely need it. Hybrid routing is the default pattern now, not an optimisation.

5. Multimodal and embodied reasoning crossed into usable

Vision-language models that act — read a screen, operate an interface, reason about a physical scene — are now good enough to be deployed for narrow tasks. Google's Gemini Robotics ER 2, released in July 2026, is explicitly positioned as an “embodied reasoning” model for machines that need to understand their surroundings and work through multi-step physical tasks. Source ↗

Two consequences worth planning for: visual regression testing and UI automation become genuinely automatable (the agent looks at the screen and judges, instead of asserting on selectors); and robotics prototyping gets a reasoning layer that used to require a research team. Both are still narrow, both still need a human gate on anything irreversible.

6. Agent security became its own discipline

Two years of deployments produced a clear lesson: prompt injection is not a prompt problem. If an agent reads external content and holds tools, that content is an input to your control flow. The defences that work are architectural — least-privilege tool surfaces, egress restrictions, allowlists, sandboxing, and monitoring that can stop a run mid-trajectory rather than grading it afterwards.

Expect this to be where procurement scrutiny lands in 2027: not "which model", but "show me the tool surface, the gate, and the trace".

7. What this changes for enterprise and SAP work

  • Joule-style assistants stop being the only option. With capable local models, an agent over your own ABAP and process documentation no longer requires sending code to a third party — which unlocks the cases that were blocked last year.
  • The migration use case is real and narrow. Classification and triage of findings, test-script drafting, reconciliation worklists. Not "AI migrates your system".
  • Cost control moves to the architecture layer. Routing, caching, reasoning depth and model size are now the four levers, and they belong in the design review.
  • Evaluation is the asset. Your eval set, your traces and your domain data are the durable advantage — models are a commodity you re-select every quarter.

What I am watching next

  1. Trajectory-level evaluation as a product category — the tooling gap is obvious and someone will fill it.
  2. Agent identity and permissions in enterprise directories: agents as first-class principals, with their own credentials, budgets and audit trail.
  3. On-device multimodal at 5 W — the point where a drone or a phone runs a vision-language model locally changes what "autonomous" means.
  4. Whether the reasoning-dial approach survives contact with regulators, who will want to know what the model was thinking and at what depth.

Sources and how this page is maintained

This is a rolling page, updated when something changes a delivery decision — not on a schedule. Claims carry the date they were published; a link is given where the primary source is public. Where I could not verify a number, it is not stated.

Building something where one of these shifts matters? Tell me the constraint — I answer with what I have actually run, not what benchmark it scored.