Alibaba just built a chip for AI agents. What it means for the UK skills gap.

AI News dropped a story this week that most people will read as a chip story. It isn’t. It’s a skills story dressed up as a chip story.

The headline: Alibaba’s semiconductor subsidiary T-Head has launched the Zhenwu M890, a new AI processor that triples the performance of its predecessor. Three-times speed is interesting but not unprecedented. The unprecedented bit is what the M890 is for: it’s purpose-built for AI agents — software systems that retain long stretches of context, coordinate with other models in real time, and execute multi-step tasks with limited human intervention. Those workloads put very different demands on memory bandwidth and inter-model communication than the inference chips most of the market is still optimising for.

Alongside the chip, Alibaba shipped Qwen 3.7-Max, its newest flagship large language model. The model’s headline spec: it can run continuously for up to 35 hours without performance degradation. That number only makes sense if the design target is sustained autonomous operation, not single-turn chat.

And the roadmap: V900 in Q3 2027, J900 in Q3 2028, each expected to deliver another roughly threefold gain. This isn’t a one-off product. It’s a sustained, deliberate cadence aimed at owning the agentic compute stack.

The skills story underneath the chip story

Chip vendors only invest in silicon for workloads they expect to dominate. T-Head shipped 560,000 Zhenwu units to 400+ external customers across 20 industries before the M890 launch — that’s real production volume telling Alibaba which workloads its customers actually run. They’ve decided the answer is agents.

Same pattern at Huawei (Ascend roadmap), at Nvidia (Vera platform), at every major cloud provider. The hyperscalers are pricing forward what they think the next 36 months look like. That picture is agents, not chat.

For UK businesses, the implication is uncomfortable. The compute will be there. The models will be there. The skill of designing, building, evaluating and governing AI agents is not — and the gap is widening every month a team doesn’t close it.

We have UK employers spending serious money on Copilot, Claude and Gemini licences whose teams can’t build a single agentic workflow end to end. The licence is the easy part. The skill is the bottleneck. And the gap is no longer ‘teach the team to write better prompts’ — it’s ‘teach the team to design systems that coordinate models and tools without a human in the loop.’ — Rod Doyle, Director, TESS Group

What “AI agent skills” actually means in 2026

An AI agent isn’t a chatbot with extra steps. The skill set is genuinely different. A team that can ship and run agents needs competence across at least five areas:

1. Agent design. Decomposing a job into the right number of agents with the right scope; picking which model runs which step; defining the success criteria the system optimises for.
2. Tool integration. Wiring agents into source systems with the right permissions, error-handling and rollback paths. Microsoft Copilot Studio, Google Gemini Gems, Make.com, Zapier, custom API connectors.
3. Multi-agent orchestration. Patterns for coordinating multiple agents on a shared workflow — sequential, parallel, supervisor/worker, RAG-augmented. Knowing when to use which.
4. Evaluation at scale. When one Spark job overnight could produce 100 outputs you didn’t read, you need sampling, automated checks, calibrated trust thresholds, and human-in-the-loop escalation rules.
5. Governance and audit. Who’s responsible when an agent acts? What’s the audit trail? How do you satisfy UK GDPR, FCA Consumer Duty, the new EU AI Act deadlines?

Most teams are at level zero on items 2–5. That’s the skills gap chip vendors are building around.

How TESS trains AI agent skills today

Three routes, depending on time horizon and depth.

1. The flagship: AI & Automation Practitioner Level 4 apprenticeship (ST1512)

Our ST1512 apprenticeship is built specifically around agentic AI as a core competency, not an afterthought. Learners design and ship AI agents using Microsoft Copilot Studio, Google Gemini Gems, AppSheet, Power Automate, Make.com and Zapier — covering single-agent automations, multi-agent workflows, and agentic automation patterns including RAG. 15 months (deliverable in 12), no coding required, 100% funded for UK SMEs under £3m payroll. Embedded qualifications include Microsoft AI Business Professional (AB-730) and NCFE Level 3 in AI Prompt Mastery, Data and Cyber Security Practices.

This is the person who’ll actually run your agent stack 12 months from now — the workflow designer, the orchestration owner, the internal trainer.

2. The strategic layer: Level 5 AI Leadership Units (AU0009 / AU0010 / AU0011)

For the director making the call on which agents the business deploys, with what oversight. Three short modular units, each 30 hours, £750 each, 100% funded for SMEs under Funding Model 99:

• AU0009 AI Strategy — deciding where agents fit in the operating model
• AU0010 AI Adoption & Governance — the audit trail, risk register, vendor due diligence, regulatory mapping
• AU0011 AI Delivery & Transformation — running the actual rollout

3. Whole-team baseline: AI workshops

Our existing AI workshops — Microsoft Copilot, Google Gemini, AI for Leaders, Building AI-Ready Teams, Automating Workflows — cover baseline agent literacy at half-day to two-day length. Tool-agnostic where useful, deeply tool-specific where the value is in the buttons.

What’s missing: a dedicated “Build AI Agents” short course

We’ve had the same conversation with five UK employers in the last fortnight. They don’t want a 15-month apprenticeship for the immediate need. They don’t want a generic prompt-engineering workshop. They want a focused 2-day intensive on how to design, build and ship a working AI agent using the tools they already have — Copilot Studio, Gemini Gems, Make.com, Zapier — with patterns they can take back to the office and apply on Monday.

That course doesn’t exist in the TESS catalogue today. The closest is our Automating Workflows short course (great for classic process automation but pre-dates the agentic shift) and our Generative AI Bootcamp (broader scope, includes agents as a section). We’re actively scoping a dedicated “Build AI Agents” 2-day short course to fill the gap, alongside the apprenticeship route.

What to do this week

Three things, in order:

1. Map your agentic surface area. Which workflows in your business genuinely need an agent vs a simple automation vs a human? Write it down. Most teams discover the list is shorter and more focused than they expected.
2. Pick one person who’ll own the agent stack and put them on ST1512. They’ll be the workflow designer, the orchestration owner, the internal trainer. £0 to the business if you’re an SME under £3m payroll.
3. Book a 25-minute discovery call. We’ll match your team profile to the right combination of apprenticeship, units and workshops — and put you on the “Build AI Agents” short-course pilot list if it fits.