6 · Market Intelligence

AI architecture needs a technology radar inside the spec, because the vendor/model/tool landscape shifts constantly. This is knowledge, not architecture — but knowledge that drives architecture.

Landscape — an AI-native technology radar. Modeled on Thoughtworks' Tech Radar, with an extra blocked ring for hard constraints. Dated and scoped, so it expires on schedule.

landscape: Coding Harnesses
  as_of: 2026-06-02
  review_by: 2026-06-16
  scope: Hermes

  adopt:
    - Codex
      reason: best current cost/performance for Hermes backend use
  trial:
    - GLM
      reason: cheapest credible coding-agent substitute
  assess:
    - Opus
      reason: highest cognition, but expensive
  caution:
    - Qwen Coder
      reason: weak MCP connection reliability in current tests
  blocked:
    - Claude Code
      reason: current interpretation of Anthropic usage rules blocks backend harness

Because it is structured rather than prose, it becomes queryable: Why not Claude Code? What is the cheapest substitute? What needs retesting next week?

Capability matrix — the most-used view, made test-backed. Every matrix carries provenance (as_of, review_by, scope, evidence, method, confidence).

matrix: Coding Agent Candidates
  as_of: 2026-06-02
  scale: 1..10
  higher_is_better: true
  evidence: [bench#CODE-HARNESS-2026-06]

  candidate     coding  mcp  cost_efficiency  cognition  backend_allowed
  Claude Code   9       10   5                9          no
  Codex         9       7    8                8          yes
  GLM           8       6    10               8          yes
  Opus          9       9    4                10         yes
  Qwen Coder    7       4    9                7          yes

Required metadata on every matrix: as_of: review_by: scope: evidence: method: confidence:

Tradeoffs — the shape of each option:

tradeoff: Claude Opus
  + strongest reasoning
  + strongest MCP
  - expensive

tradeoff: GLM
  + cheapest
  + good coding
  - China-centric

tradeoff: Codex
  + good coding
  + OpenAI ecosystem
  - weaker MCP ecosystem

Recommendations — answer "what should I choose?":

recommendation:
  if: prioritize = MCP      -> choose: Claude Code
  if: prioritize = Cost     -> choose: GLM
  if: prioritize = Harness  -> choose: Codex