Technology · 12 Technology Radar 2026
A working radar, not a trend list. Each entry says which ring it sits in, why it moved, and what the choice will cost a client three years from now. Updated September 2026 by the team that builds and maintains the work.
Updated September 2026 You own everything Reply within one business day
Forecast engine
Demand forecast - next 8 weeks
94.2%
Accuracy
+18%
Vs baseline
8w
Horizon
Retrained nightly on 14 months of order behaviour.
OlDevs keeps a technology radar with four rings.Adopt for defaults on new work, Trial for tools proven on at least one client project, Assess for things we prototype internally, and Hold for choices we no longer recommend. The Vancouver studio reviews it every quarter and moves an entry only when a real project has tested it. Every recommendation is weighed against who will maintain the system and what changing course later would cost.
Key facts
12 · Technology Radar 2026
These are the radar entries that changed a default or a recommendation since our last update, with the ring each one now sits in.
Right for every project we now run. Agents such as Claude Code and GitHub Copilot draft code, tests and migrations inside a reviewed pull request flow, and an engineer owns each merge. The later cost is discipline: without tests and conventions, agent output builds technical debt faster than a person could.
The right choice when an assistant or agent must read a CRM, ticketing system or database. The major model providers and IDEs now support MCP, so one server serves many clients. The later cost is governance: each server needs authentication, scoped permissions and logging, or it becomes an open door to your data.
Right when a workflow has more than one step and a clear success measure, such as triaging support tickets or matching invoices. We pair LangGraph or a vendor agent SDK with an evaluation suite that scores every change before release. Skip the suite and you pay later in silent regressions nobody can reproduce.
Right when staff or customers need answers grounded in your own documents. RAG has matured: pgvector inside PostgreSQL, hybrid search and re-ranking cover most cases without a separate vector database. The ongoing cost is content hygiene: stale or duplicate documents degrade answers, so someone must own the corpus.
Right for private, offline or low-latency features on recent phones: summaries, form filling, classification. Apple's Foundation Models framework and Gemini Nano through ML Kit make this practical, but capability varies by device. The cost is a fallback path to a server model for older hardware and harder requests.
Both are excellent for tooling, scripts and edge functions, and Deno 2's Node compatibility is real. We still hold them as the main runtime for long-lived client systems because hosting support, monitoring and hiring pools favour Node.js LTS. Choosing them today can mean fewer people to maintain the system in 2029.
Stack
Concrete tools, current at the time of writing; confirmed for your project at scoping.
AI delivery and agents
Front end
Runtimes and languages
Data and identity
Quality, performance and compliance
What changed
What moved in this area and what it means for your build.
Sep 2026
Coding agents graduated from Trial after a full year across client projects. For clients this changes how estimates are built: less time drafting code, more time on review, tests and acceptance.
2026
The first compliance date for state and local government web content under the US Department of Justice rule fell in April 2026, with smaller entities following in 2027. Canadian vendors selling to those bodies inherit the requirement; WCAG 2.2 AA covers it.
2026
Operating systems now create, sync and share passkeys by default, and the major identity providers support them. Passwordless sign-in is now a product expectation rather than a differentiator, and a password-only login reads as dated.
Late 2025
Node.js 24 entered long-term support, Python 3.14 shipped with an officially supported free-threaded build, PHP 8.5 arrived, and PostgreSQL 18 introduced asynchronous I/O. New builds start on these versions; existing systems get a dated upgrade plan.
2025
Apple's Liquid Glass design language and Android's Material 3 Expressive changed how native apps are expected to look and move, and iOS 26 opened the on-device Foundation Models framework to developers. Apps that skip the refresh look older faster.
2025
Tailwind v4 moved configuration into CSS, Svelte 5 runes became the standard reactivity model, Vite began its move to the Rolldown bundler, and Microsoft previewed a native Go-based TypeScript compiler. Faster builds, but each migration needs a planned window.
How we choose
We start with the team who will own the system after launch. If a client's developers know Laravel, PHP 8.5 beats a Node rewrite, however much we like the alternative.
Nothing reaches Adopt on the strength of a blog post. An entry advances only after it has shipped in a client project or a paid pilot and survived a maintenance cycle.
Canadian public sector and health clients often need data kept in Canada. That rules out some managed AI and database services and limits which ring a tool can reach.
We estimate what it costs to leave a technology, not only to adopt it. Open standards, exportable data and large hiring pools score higher than lock-in.
WCAG 2.2 AA, Quebec Law 25 and ADA Title II shape our defaults: accessible components, consent and data-request flows, and audit logging are in from day one.
We prefer versions with a stated support window, such as Node.js LTS releases and PostgreSQL's five-year policy, so a client knows when the next upgrade lands.
Who it's for
Whether the audience is a customer, a member, a citizen, or your own team, the choice has to hold up under real use.
You have existing systems and internal IT. The radar shows which of your platforms are still in Adopt, which need an upgrade plan, and where AI agents can be added without replacing what works.
Accessibility deadlines and privacy law drive your priorities. Our defaults meet WCAG 2.2 AA, keep data in Canada where required, and avoid lock-in that a future procurement cannot unwind.
Many locations, one codebase. We favour platforms that scale by configuration, passkey sign-in that cuts password support calls, and on-device features that run on the phones your staff already carry.
Speed matters, but so does the next hire. We pick a boring, well-supported core, use coding agents to move quickly, and reserve Trial and Assess tools for the one feature that sets you apart.
Related technology
React, Next.js, Vue, Svelte, Angular, Astro and WordPress each earn a place on our shortlist for different…
Explore10Every OlDevs project ships with threat modelling, OWASP ASVS checks, encrypted data, single sign-on and audit…
Explore07Most AI projects never need a training cluster. We match each workload to the right silicon, rent before we…
ExploreFAQ
Yes, on every project since 2025. Agents work inside our repositories under the same rules as a junior developer: they open pull requests, the tests run, and an engineer approves every merge. We use enterprise accounts with data retention switched off, so client code never trains a public model. You own everything the agents produce.
Stay on Node.js LTS for the core of a long-lived system. Bun and Deno 2 are fast and pleasant, and we use them for scripts, tooling and edge functions where they fit. They sit in Hold as a primary runtime not because of quality but because of the smaller pool of hosts, monitoring tools and developers around them, which your team will feel in year three.
For content-heavy sites and applications that mix public and signed-in pages, yes; Server Components cut the JavaScript sent to the browser and Next.js is mature enough to be a safe default. If the product is a small, highly interactive app and your team prefers a simpler mental model, Svelte 5 with SvelteKit is in Trial and has done well.
Interaction to Next Paint measures how quickly a page responds after a tap or click, and it replaced First Input Delay in Google's ranking signals. Sites built with heavy client-side JavaScript often passed the old metric and fail this one. We measure INP on real devices, trim third-party scripts, move work off the main thread, and serve over HTTP/3.
Almost never. Blackwell-class hardware is widely available from cloud providers, and renting it by the hour for training or batch inference is far simpler than owning it. We recommend owning hardware only when data cannot leave your premises or a model runs continuously at a scale where rental becomes the dearer option, and we model that before you buy.
Directly, no; the rule binds US state and local governments and the bodies they fund. Indirectly, yes if you sell software or content to those bodies, since they will require conformance from vendors. Our standard is WCAG 2.2 AA in English and French, which exceeds the rule's WCAG 2.1 AA baseline and also satisfies Canadian federal and provincial requirements.
If you collect personal information from people in Quebec, the law applies wherever you are based. In practice that means a named privacy officer, consent that is granular and refusable, a privacy impact assessment before new data uses, data portability on request, and breach notification. We build the consent, export and deletion flows into the product itself.
Let’s connect
Tell us what you’re building. We’ll reply within one business day with recommended platforms, structure and a tailored quote — no obligation.