Weekly signal

This briefing covers the week 2026-09-28 through 2026-10-06 and highlights a concentrated set of agentic-AI releases and deployments that push agents from experiment to operational collaboration: platform-hosted always-on agents and workspace surfaces (OpenAI), reusable/versioned agent configuration for teams (Perplexity), vendor business pushes for integrated agents (Meta), model refreshes tuned for agent work (Anthropic), and a U.S. federal AI front-door rollout (White House). These moves collectively make human–AI handoff, governance, and shared context the immediate engineering and product priorities.

What changed

1. OpenAI’s DevDay (Sep 29, 2026) introduced Dots (always-on, personal/team agents), the Agents API with "computer use" (agents that can operate software UIs), collaborative Pages/Spaces for humans+agents, and new developer primitives for multi-agent workflows and hosted execution. This is a major push to treat agents as persistent teammates in product workflows rather than single-turn assistants.

2. Perplexity updated its Agent API (Sep 28, 2026) with reusable, versioned Profiles and Skills plus managed connectors so teams can centrally configure, share and update agent behavior, tools, and credentials across apps. That materially reduces per-project setup and helps apply consistent policies across agent deployments.

3. Meta announced Meta Enterprise Platform (Sep 28) and expanded Muse for Small Business (Sep 29), positioning Muse as a connector-rich, always-on agent for business workflows (ads, shopfronts, bookkeeping connectors). This signals major vendor investment in agent-to-business integrations and marketplace-style connectors.

4. Anthropic shipped Claude Sonnet 5.5 (Sep 28), a lower-cost, faster model in the Sonnet line optimized for well-scoped agent tasks (documents, slides, routine coding) and made broadly available across clouds — an example of model-tiering for different agent roles.

5. The U.S. administration launched an America.gov front door and published an executive order (Sep 29) establishing a unified AI-powered portal for federal services. This is one of the earliest large-scale public-sector attempts to place agentic interfaces as the primary citizen interaction surface.

What to do with it

• Treat agents as workflows, not chatboxes: design goals, approval gates, and escalation points up-front; instrument iteration cost and human effort as success metrics.
• Version and share agent configurations (profiles/skills). Adopt reusable, centrally managed connectors and apply change-control to agent instructions and secrets.
• Add human-in-the-loop gates and LLM-as-judge style checks for high-impact actions; map out approval UX and audit trails before connecting agents to critical systems.
• Re-evaluate model tiers for different agent responsibilities (Sonnet-like for fast routine work; Opus/Astra-like for complex judgment) and budget accordingly.
• If you integrate with public services (U.S. or other governments), plan for citation, escalation, and dispute-handling flows; expect audits and rapid policy changes.