May 2026Agentic AI7 min read

Agentic AI is having its moment. Here's what's real and what's not.

Agentic AI - models that do things instead of just chatting - is the loudest topic in software right now. Some of the noise is real. A meaningful chunk is not. Here's our honest read on what's landed, what hasn't, and the five questions worth asking before you write a cheque.

What's actually shifted in the last 12 months

agent framework downloads

growth across LangGraph, AutoGen, Pydantic AI (2025 vs 2024)

70%+

of new dev hires

report using a code agent daily (Stack Overflow 2025 survey)

$1B+

agent-startup funding

in 2025 alone, per Crunchbase tracking

The numbers above are directional, not gospel - different trackers count things differently and the framework figures include a lot of toy projects. But the trend is real and consistent across every credible source: agent tooling is the fastest-growing category in the AI stack, and code agents in particular have crossed from novelty to default tool inside software teams.

Where it's genuinely working

Code generation

Cursor, Claude Code, GitHub Copilot Workspace, Devin and Cognition's spin-offs. Real productivity, real shipped code. Tests are the safety net.

Customer support triage

Routing, drafting first replies, escalating the rest. Klarna, Intercom Fin, and a hundred quieter deployments showing 30-50% deflection rates.

Internal ops

Calendar wrangling, expense triage, report drafting, inbox triage. Saves hours per person per week. Boring, profitable.

QA + testing

Agents that explore your app, generate test cases, report regressions. The category that quietly delivered in 2025.

Where it's overhyped

Autonomous money movement

Demo-friendly, deployment-hostile. The cost of a single wrong action exceeds a year's saved time. Use a human gate.

“Replace your sales team”

Trust and tribal knowledge are the job. Agents help with research and admin, not the relationship.

Fully-autonomous companies

Marketing copy. The closer you look, the more humans you find. The interesting case is augmented teams, not replaced ones.

Long-running self-supervised agents

Agents that run for hours unsupervised drift. The further from human oversight, the worse they get. Short loops with check-ins still win.

The pattern is consistent. Agents excel where the work is bounded, reversible, and has a fast feedback loop. They struggle where the work is open-ended, irreversible, or depends on relationships and context that aren't written down.

The five questions before you build (or buy) one

What is the worst single action this agent can take? If the answer involves money leaving the business, customer trust, or anything legal - put a human in front of it.
What does “done” look like? Agents need a stop condition. If you can't define it, you don't have a project, you have a research grant.
What's the cost-per-loop? Multiply tokens-per-call by calls-per-task by tasks-per-day. Now multiply by 30 days. If that number scares you, the unit economics don't work.
How will you tell when it goes wrong? Agents fail silently. You need logging, human spot-checks, and an honest review cadence built in from day one.
What replaces this if the API goes down? Frontier APIs do go down. The fallback plan matters more than the happy-path plan.

What we expect next

Tool-use APIs converging across vendors. The Model Context Protocol (MCP) and equivalent open standards are starting to let the same agent talk to any model.
Memory becoming a first-class primitive. The 2025 trend was longer context windows; the 2026 trend is structured memory that persists across sessions without bloating the prompt.
Multi-agent setups stabilising. Specialised agents that hand off to each other (a researcher, a writer, a reviewer) are starting to outperform single-agent loops on complex tasks.
The compliance story catching up. Expect EU AI Act guidance, UK regulator templates, and insurer questionnaires that assume you can describe what your agents do, who supervises them, and how you log it.

Where to go next

What is an AI agent? - 5-minute primer - the plain-English definition with archetypes.
The 60-minute AI readiness self-assessment - score whether your team and data are actually ready.
Context windows, hallucinations and agents - the three concepts that filter most snake oil.

Agents will reshape the operating model of small businesses over the next two years - but only the boring, well-bounded ones. The autonomous-everything pitch is still pitch. The useful question is not can we have an agent? It is which one repeating task in the business is bounded enough that a tireless junior assistant would crush it?