Abbeal Tech Radar 2026: what we adopt

Methodology

Abbeal's Tech Radar is inspired by the ThoughtWorks matrix, adapted to our tri-geo context. Every three months, Tech Leads from the three hubs (Paris, Montréal, Tokyo) vote on the key technologies of our missions. Four levels:

• Adopt — validated in production on at least 3 missions, performant, maintainable. Default mandate on new missions.
• Trial — positive validation on 1-2 client missions. Recommended if project context fits. Escape hatch required.
• Assess — promising but not yet tested in prod. To experiment with in POC or side project. Not in critical client missions.
• Hold — tech we're actively retiring, or advising against starting with. With documented rationale.

The vote is not democratic: each Tech Lead brings measurable data (latency, memory, time-to-merge, bugs, learning curve). Choices are documented in a Notion file shared with clients under mission.

The current radar, technology by technology

Read straight from Tech Radar Q4 2026, in editorial order. Each line gives the technology and why it sits where it does.

Adopt — 14 technologies

• Rust (systems + robotics) — Zero-cost abstractions and memory safety without a GC's runtime cost. 4 production releases in Q2 2026 on our robotics systems and edge proxies. Mature Tokio / Axum / sqlx ecosystem; latest stable: 1.98.
• TypeScript strict — strict + noUncheckedIndexedAccess by default on every new engagement. TypeScript 7, the native port released in July, compiles 8 to 12 times faster according to Microsoft, with the same type semantics: adopt it as soon as the project's tooling supports it.
• Claude Opus 5 + Sonnet 5 (API) — Our default models for AI-first engineering: Opus 5 for code and complex agentic work, Sonnet 5 for speed and cost. Prompt caching and structured tool use in production. The 4.x generation is still available but no longer our recommendation.
• Claude Code (Anthropic) — Our default AI development tool since May 2026, internally and on client engagements. Used with explicit guardrails: scoped permissions, sandboxing, human review before merge.
• Production RAG (embeddings + hybrid search) — Retrieval + generation with a vector database and reranking. Multilingual FR/JA field-tested with 4 clients. Every pipeline ships with its evaluation set (see “Continuous LLM evaluations”).
• dbt (Data Build Tool) — Versioned, tested and documented SQL: our transformation standard, deployed at 4 data clients in 2026. dbt Core 2.0, released in September on the Fusion engine, changes the metadata artifacts: plan the migration project by project.
• Turso libSQL (edge SQLite) — Distributed, edge-native SQLite. libSQL is what runs Turso Cloud in production, and our Payload CMS since June 2026. The Rust rewrite (Turso Database) is still in beta: not the same thing.
• OpenTofu — Systematic migration from Terraform on new projects and renewals, with a tested and documented state migration. Current branch: 1.12, released in May.
• SecNumCloud V3.2 (ANSSI) — ANSSI's sovereign cloud framework, made applicable by the decree of August 12, 2026 to French State services, State operators and public interest groups for their sensitive data. We support migrations from hyperscalers to qualified offerings (OVHcloud, 3DS Outscale, NumSpot).
• Next.js 16 — Our default frontend stack, with SSR / ISR. Version 16.3 (August 2026) sharply cuts dev-server memory and speeds up builds. Abbeal.com runs on it. Cache Components now has its own entry, in Trial.
• ROS 2 + Nav2 (production) — The modern robotics standard: DDS transport, real time, native multi-robot. Nav2 in production at 2 industrial clients. Lyrical Luth (LTS, May 2026) for new projects; Humble reaches end of support in May 2027, so start migrations now.
• Vercel + Fluid Compute — Default deployment platform for our frontends and APIs: standard Node.js runtime, fast ISR revalidation. Fluid Compute is the default for new projects, with a 300 s default execution time since June.
• pgvector — PostgreSQL as a vector database: our default for RAG under 5 million vectors. No extra infrastructure when Postgres is already in production. Current version: 0.8.6.
• Agent guardrails (sandboxing, least privilege) — Every agent that touches code or data runs in a sandbox, with only the permissions it needs and a trace of every action. It is the precondition for every AI entry on this radar: no guardrails, no agent on a client engagement.

Trial — 10 technologies

• Claude Agent SDK — Claude Code's engine as a library (TypeScript and Python), to build our own agents. Tested on code review and feature generation pipelines. Promising, with a real learning curve.
• LiteLLM router — Multi-provider abstraction to switch between Claude, GPT and Gemini without rewriting. Two malicious versions were published on PyPI on March 24, 2026: we run it with pinned, hash-verified versions, keys kept outside the proxy, and a dependency inventory.
• Pinecone (managed vector DB) — In production at 3 RAG clients above 10 million vectors. Expensive but with a solid SLA. Below 5 million, pgvector is enough.
• DORA metrics automation — Lead time, deployment frequency, MTTR and change failure rate, instrumented through Sentry and GitHub Actions. More useful than ever as agents write a growing share of the code: it is the measure that tells whether it holds up.
• MACH architecture (composable commerce) — Microservices, API-first, cloud-native, headless. Delivered in luxury retail. Beware over-splitting below 200,000 SKUs: integration cost quickly outweighs the gain.
• Cloudflare Workers + Durable Objects — Distributed edge compute for low global latency. Used for AI proxies, rate limiters and geo-distributed WebSockets.
• Edge AI (NVIDIA Jetson Orin / Thor) — AI inference at the edge on mobile robots and industrial kiosks, deployed on 3 sites in Q2 2026 with Orin. Jetson Thor, available since August 2025, delivers 7.5 times Orin's AI compute according to NVIDIA: worth evaluating for new projects.
• Cache Components + Partial Prefetching (Next.js) — Next.js's new explicit caching model (“use cache”), with near-instant navigations. Stable, but still opt-in in 16.3 and set to become the default: we roll it out project by project, not yet by default.
• MCP (Model Context Protocol) — The open protocol that connects agents to tools and data. Handed to the Agentic AI Foundation (Linux Foundation) in December 2025 and supported by the major AI clients. We already use it; every MCP server is treated as a dependency to audit.
• Continuous LLM evaluations — A versioned evaluation set for every prompt, agent or RAG pipeline, replayed in CI whenever the model or prompt changes. It is what lets us switch models without silent regressions.

Assess — 7 technologies

• Multi-tool agents — Agent workflows with more than 3 tools. Reliability is improving, but long-run cost and attack surface grow with permissions: we don't move them forward without the guardrails listed in Adopt.
• Rust web frameworks (Axum, Loco) — Axum is stable; Loco, the Rails-like framework, has reached 1.x. Promising for performance-critical APIs, but not yet our default choice.
• WASM component model — WASI 0.3, released in June, brings native async to components. Promising for sandboxed plugins, but tooling is still stabilizing.
• Qdrant + Weaviate (open-source vector DBs) — Open-source alternatives to Pinecone: Qdrant for simplicity, Weaviate for hybrid search at scale. Evaluated where data sovereignty requires self-hosting.
• DSPy (Stanford) — Framework to compile prompts systematically instead of hand-writing them. A LangChain alternative for deterministic LLM pipelines. Serious learning curve.
• Foxglove / Lichtblick (ROS 2 visualization) — ROS 2 and rosbag visualization, a good complement to RViz2. Foxglove is now commercial only, its open-source repository archived since 2024; Lichtblick is the active open-source fork. The choice depends on the license the client accepts.
• Claude Fable 5.1 (Anthropic) — Anthropic's most capable widely released model (September 1, 2026). Out of Hold: access, suspended from June 12 to July 1 under US export controls, has been restored. We are evaluating it on long agentic work; it can decline requests, costs twice as much as Opus 5 and requires 30-day data retention. Mythos 5.1 remains limited to Project Glasswing.

Hold — 4 technologies

• Low-code frameworks (Webflow, Bubble) — Fine for fast prototyping. Beyond 6 months in production, technical debt is inevitable, and clients often ask us to migrate them to React / Next.js.
• B2B crypto payments (Europe / Japan) — Client volumes in France and Japan remain marginal compared with SEPA and Stripe, despite an EU framework (MiCA) fully applicable since late 2024. No tangible B2B traction in 2026; to be reassessed in Q1 2027.
• jQuery + legacy JS (Backbone, Ember) — The Hold targets legacy applications, not the library: jQuery is maintained (4.0 released in January 2026). The cost lies in the old stacks around it; we migrate to Next.js, Astro or Svelte depending on the need.
• AWS Lambda + Node 18 / 20 — Node 18 and Node 20 are deprecated on Lambda; function updates will be blocked on March 3, 2027. Target: Node 24, supported until April 2028, rather than Node 22 (deprecated in April 2027). Move to arm64 (Graviton): up to 34% better price-performance according to AWS.

How we use it on client missions

On every new Abbeal mission, the Tech Radar is part of the technical mandate negotiated at kick-off:

1. Default mandate: mission stack composed of Adopt techs unless explicit client agreement.
2. Escape hatch: if client wants to deviate from Adopt (e.g. keep an AWS Lambda function on Node 18 or a jQuery front end), we document the derogation with justification + eventual migration plan.
3. Quarterly review: every 3 months, we review with the client whether Trial techs we've deployed move to Adopt on Abbeal side, and whether the contract should adapt.

This discipline lets a client engaging Abbeal on a 12-24 month mission guarantee they won't be stuck on an obsolete tech by the end.

All editions

• Tech Radar Q4 2026 — published September 18, 2026, 35 technologies
• Tech Radar Q3 2026 — published July 8, 2026, 31 technologies
• Tech Radar Q2 2026 — published April 15, 2026, 11 technologies

Tech Radar editions index