Agents as Infrastructure: harnesses, context, and cost controls

Anthropic’s Claude Tag update lets Slack agent read full conversations and jump in unprompted. Anthropic updates Claude Tag so its Slack agent can ingest entire threads and proactively interject without explicit prompts. Outcome engineers must design permission, intent, and provenance controls for agents that now act as active participants in collaboration platforms.

Google brings Antigravity under Gemini Enterprise for granular spend controls. Google centralizes Antigravity billing, adding pooled quotas, granular spend caps, overage controls, and usage metrics to Gemini Enterprise. Build quota enforcement and cost telemetry into your orchestration layer so agents can run productively without creating surprise bills or denial-of-service by cost.

The Harness Is the Company. Shrivu argues companies must become harnesses around models, shifting humans to provide context, judgment, and review while agents execute core work. Treat the harness as product: define human review gates, audit trails, and clear owner workflows so agent output becomes a repeatable, auditable artifact.

Agent Is Not the Model. The post clarifies that agents are orchestration and stateful harness layers distinct from the underlying inference models and services. Decouple agent logic, state management, and orchestration from model selection so you can iterate harnesses, swap models, and validate outcomes without retraining or redeploying the whole stack.

My agent.md to improve LLM-assisted code quality. Fabien Sanglard publishes agent.md, a standardized coding-style prompt spec that makes IDE agents produce consistent, production-ready code. Version and enforce agent.md-like artifacts in CI to keep agent outputs legible, reduce manual corrections, and preserve team coding standards.