Skip to main content

tech radar · v2 · november 2026

What I'm betting on, and what I'm not

Years ago I launched my first company's Tech Radar to give engineering teams a shared formulary for what to adopt, trial, assess, and hold. This is the personal version. 102 blog posts written since May 2025, distilled into the bets I am willing to defend out loud.

last reviewed November 4, 2026 · click any blip or any item below for the argument behind it

ADOPTTRIALASSESSHOLDEngineering PracticesAI Tools & ArchitectureLeadership PosturesIndustry Narratives12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697
Adopt — use today. default in new workTrial — in production. convinced of directionAssess — watching. not yet bet onHold — refuse, or warn against

quadrant 01

Engineering Practices

How software actually gets built, reviewed, and shipped in an agentic SDLC.

Adopt

Use today. Default in new work.

  • 01Spec-driven development
  • 02Standards in files, not heads
  • 03Architecture as agent guardrails
  • 04Pipeline-first delivery
  • 05Verification-first development
  • 06Fail to unknown, never to green
  • 07Conformance suite as executable oracle
  • 08Count the work for everyone
  • 09Name an owner for every agent
  • 10Separation of duties for agents
  • 11Context Spine / Intent Stack
  • 12Enforced design systems, not voluntary ones

Trial

In production. Convinced of direction.

  • 13Answer-key-first modernization
  • 14Operability in the spec
  • 15Spec review as the new code review
  • 16Multi-agent specialist workflows
  • 17Prompt review as quality gate
  • 18Workflow inversion over agent loops
  • 19Walk one change backward
  • 20Two lanes for dependency updates

Assess

Watching. Not yet bet on.

  • 21Dark factory model for software
  • 22Agent workflows as versioned releases
  • 23Test enforcement separately from refusal

Hold

Refuse, or warn against.

  • 24Vibe coding without specs
  • 25Kubernetes for one service
  • 26Agent teams as a default
  • 27Instructions as access control
  • 28Line-by-line review as primary quality bar

quadrant 02

AI Tools & Architecture

The seam between local and frontier, the patterns that compound, the ones that do not.

Adopt

Use today. Default in new work.

  • 29Claude Code as build system
  • 30Hybrid local + frontier inference
  • 31MCP as integration substrate
  • 32AI providers like cloud providers
  • 33Design for sovereign model risk
  • 34Own the learning loop, not the model
  • 35Opinionated defaults, explicit escape hatches

Trial

In production. Convinced of direction.

  • 36Open-weight models for regulated workloads
  • 37Personal agents for everyday follow-up
  • 38Custom-trained frontier models (Nova Forge style)
  • 39Design the platform for the agent persona
  • 40Context as a platform capability
  • 41Invisible AI over the portal chatbot
  • 42Keep the agent replaceable
  • 43LLM-consumable design systems
  • 44Staff agents by the cost of a mistake

Assess

Watching. Not yet bet on.

  • 45Intent interface over the golden path
  • 46Fully autonomous coding agents
  • 47Delegated identity for agents
  • 48AI design tools beyond Figma
  • 49Decision models for typed judgments

Hold

Refuse, or warn against.

  • 50Multi-cloud as a hedge
  • 51Prompt-whispering as a discipline

quadrant 03

Leadership Postures

How engineering leaders should show up, what they should reward, and what they should refuse.

Adopt

Use today. Default in new work.

  • 52Stay close to the craft
  • 53Principles over personalities
  • 54Accountability flows upward
  • 55Clarity beats consensus
  • 56Turn complexity into leverage
  • 57Servant leadership in tech
  • 58Invest in framers, not finishers
  • 59Judgment as the scarce resource
  • 60Bargain honestly with disruption
  • 61Value per verified outcome

Trial

In production. Convinced of direction.

  • 62FinOps as a core engineering competency
  • 63Build for succession
  • 64Designed apprenticeship over accidental
  • 65The AI Operating Ledger
  • 66Spend the saved hour on purpose
  • 67The developer experience ledger
  • 68Evaluation as its own capacity

Assess

Watching. Not yet bet on.

  • 69Pairing on specs and reviews
  • 70Look past the rollup

Hold

Refuse, or warn against.

  • 71Mandate-driven AI adoption
  • 72Building monuments to yourself
  • 73Headcount-first AI reorgs
  • 74Agent time as a success metric
  • 75Personal agents as AI strategy
  • 76PowerPoint-only leadership

quadrant 04

Industry Narratives

The stories the industry tells itself. Some are true. Most are someone's quarterly earnings call dressed up in a hoodie.

Adopt

Use today. Default in new work.

  • 77"Specs beat vibes"
  • 78"The bottleneck is never the stack"
  • 79"Agents advise; humans decide"
  • 80"From toil to judgment"
  • 81"Trust the gate, not the actor"

Trial

In production. Convinced of direction.

  • 82"Trust compounds faster than velocity"
  • 83The builder’s bargain
  • 84"The frame is the bottleneck"
  • 85Open weights as commoditization strategy

Assess

Watching. Not yet bet on.

  • 86AI infrastructure capex bubble
  • 87Open-source funding under AI
  • 88Capability as a governed resource
  • 89The rewrite taboo was a price tag
  • 90Intelligence as infrastructure
  • 91Safety rules as incumbent moats
  • 92Agent access follows the margin

Hold

Refuse, or warn against.

  • 93"Workforce optimization" language
  • 94"AI is coming for your job"
  • 95"AI replaces engineers"
  • 96The salary-for-subscription swap
  • 97Vendor-as-operating-model