Agent Ops: Context, Cost Controls, Local Agents, and Hardware

Anthropic’s Claude Tag update lets Slack agent read full conversations and jump in unprompted. Anthropic updates Claude Tag so a Slack agent can read entire threads and proactively join conversations. Outcome engineers must treat proactive, multiplayer agents as first-class orchestration problems—design context limits, handoff rules, and audit trails before they start acting on behalf of users (Principle 03, 09).

Google brings Antigravity under Gemini Enterprise for granular spend controls. Google centralizes Antigravity billing with pooled quotas, per-agent spend caps, and overage controls inside Gemini Enterprise. This changes how you govern agent fleets—budgeting, quota policies, and usage metrics become core components of your outcome gatekeeping and cost-aware orchestration (Principle 10, 12).

Junie now runs entirely offline — can you spare a 64 GB M5 Mac?. JetBrains ships Junie Local, an on-device coding agent running Qwen3.6-27B for fully offline developer workflows. Local agents shift your threat model, latency, and privacy posture—build deployment patterns, CI hooks, and validation suited to offline inference and disconnected UX (Principle 07, 06).

Agent Is Not the Model. The post argues agents are harnesses and orchestration layers distinct from underlying inference models and services. Treating the agent as an integration, policy, and observability layer forces clearer contracts, modularity, and testability in your delivery pipeline—don’t conflate model research with production agent design (Principle 06, 11).

Up to 30x More Work Per Watt: NVIDIA Vera Rubin NVL72 Sets a New Efficiency Standard for AI Agents. NVIDIA claims Vera Rubin NVL72 delivers up to 30x higher agentic-workload throughput per megawatt than previous racks. That changes capacity planning and unit economics for agentic systems—revisit placement, batching, and observability to exploit much lower token costs and new scaling trade-offs (Principle 09, 12).