Agent Ops Brief: Shared memory, deploys, tracing, on‑device, ADK risk

Asana’s AI agents share memory across your company — but not your secrets. Asana turns the Work Graph into a shared, access‑controlled memory that lets coachable AI teammates operate with role-aware visibility. Outcome engineers get a concrete pattern for designing enterprise agent memory, auditability, and fine‑grained access controls (Principles 03, 06, 15).

Launch HN: Hoplite (YC S26) – Effortlessly deploy cloud coding agents. Hoplite streamlines deploying scalable, secure cloud coding agents so teams run developer agents without infrastructure hassle. That lowers the operational bar for agent orchestration and gives outcome engineers a repeatable deployment pattern for production agent fleets (Principles 09, 07).

Introducing: Cloudflare Agents. Cloudflare launches Agents and agent tracing to centralize deployment, observability, and token‑level insights across hosted agent sessions. This provides an operational blueprint for agent observability, tracing, and outcome validation you can adapt to monitor drift and enforce runtime guardrails (Principles 06, 16).

Google ADK flaws reveal what happens when AI agents trust the wrong message. Reported ADK bugs let public‑facing agents escalate privileges via agent‑to‑agent prompt injections, enabling pull‑request manipulation and credential exposure. Outcome engineers must harden agent boundaries, add CI/CD security controls, and design injection‑resistant communication to prevent privilege escalation and supply‑chain attacks (Principles 09, 15, 14).

Deploy local agents everywhere with LFM2.5-2.6B. LiquidAI releases a tiny agent model that runs tool‑using, multi‑step agents on‑device, enabling private, low‑cost agent deployments on phones and laptops. This changes architecture tradeoffs — outcome engineers can shift work to local agents for privacy and cost savings but must add sandboxing, unified telemetry, and outcome validation that spans device and cloud (Principles 07, 09, 16).