AI & Agents
Capability is no longer the bottleneck — governance of agent decisions, scaffolding design, and institutional legitimacy are.
Signals
AWS launched open-source Agent Plugins for AI coding agents with five-step cloud deployment workflow.
Constraint arbitrage — narrowing decision space before user confirms.
Kief Morris on Fowler site: humans should own loop structure, not micromanage outputs.
Practitioners publishing patterns. Within 12-18 months becomes hiring criteria.
Archgate CLI converts ADRs into machine-checkable rules for Claude Code via MCP.
Beginning of governance tooling category. Validation/provenance deficit.
Qwen3.5-35B hits 37.8% on SWE-bench Hard with one verification nudge vs Opus 40%.
Cost compression via scaffolding not scale.
Cowen argues stable AI regulation = sustainable perpetual interference.
Governance moving toward institutionalized ambiguity.
Control Surfaces
| Lever | Status | Change | Evidence |
|---|---|---|---|
| Agent loop design | Shifting | Accelerating | Fowler/Morris piece |
| Deployment defaults | Increasing vendor control | New category | AWS launch |
Watchlist
- ConfirmationMCP constraint injection adoption by 3+ teams
- InvalidationScaffolding adds 30%+ token cost
- ObservableGemini Conductor adoption rate
Falsifiers
- Frontier models remain 2-3x better
- AWS Plugins low adoption
- Anthropic ships native ADR injection
Key Unknowns
- Verify-after-edit adoption rate
- AWS Plugin override rate
- Fraction of teams with enforced institutional memory
Noise Filter
- GPT-5.3 Instant— No falsifiable claims
- Gemini Flash-Lite— Commodity trend only
Get The Signal daily
Cross-domain structural analysis, delivered every morning.