AWS released updates to its AI workspace Quick, letting users build business flows that orchestrate agents, skills and tools directly through conversation, with new deterministic steps for consistency and verifiability. Multi-agent orchestration now runs several specialized agents in parallel on a single prompt and merges their findings, while the web app keeps work running after the tab is closed.
Multi-Agent Orchestration
多智能体编排
Engineering pattern where a coordinator dispatches work to multiple independent agent sessions running in parallel and aggregate
Curated Intelligence
DeepSeek's Super-Agent concept uses a router-plus-worker two-tier architecture to enable parallel task execution and background running. Tencent Cloud's open-source Octop shipped a similar 'team mode' in v1.0.2, where a host agent decomposes tasks and expert agents execute them in parallel. The article dissects Octop's engineering implementation against the Super-Agent framework.
Anthropic has upgraded Claude Code Projects with a proactive coordinator, read-only thread sharing, scheduled reminders and branching commands, letting multiple cloud Claude Code sessions run in parallel with far less manual monitoring. The feature is in public beta for Pro and Max users on web, desktop and mobile, while Team and Enterprise are excluded.
AWS released AI-PDLC, a solution built on Amazon Quick that orchestrates five specialist agents plus an orchestrator across the product development lifecycle, with human approval embedded at key steps. It reconnects scattered signals from Slack, Salesforce, Jira and other tools to the decisions they inform, with code available on GitHub.
AFK AI says its product Comma already implements the single-Router, multi-Worker super-agent architecture later outlined by DeepSeek's Harness team, adding persistent memory, continual learning and 24/7 uptime to address the single-session context bottleneck. The system is publicly available with open-sourced code, and the piece frames human attention—capped at roughly 3-5 active sessions—as the real constraint on agent concurrency.
Perplexity launched Teammates, personal agents that coordinate Computer sessions and other agents to own long-running projects across marketing, sales, support and operations, with users able to inspect and steer work in progress. Access and features will expand over the coming weeks.
DeepSeek Harness published a design exploration arguing that the real bottleneck in multi-agent work is human attention, not model capability, and proposed a dual-layer architecture with a Router layer dispatching to isolated Worker Sessions. Open-source firm LobeHub reports running its own R&D through its 84k-star agent platform, with agents handling development and acceptance checks.
Perplexity announced Teammate, an agent positioned to carry out sustained complex work by planning and coordinating projects, adapting to shifting priorities, and submitting results for review while orchestrating Computer sessions and other Teammates. The feature requires a Perplexity Max subscription, launching first on web and later on iOS.
GitHub recorded 7.38 billion commits in Sept-2026, a 5x surge driven by agent-generated code. To handle the load, the platform expanded infrastructure and is rebuilding Git's architecture, while launching Copilot Fusion and multi-agent orchestration tools to support the emerging agentic economy.
A developer released a post-mortem on porting G-Diffuser (an F-Zero X decompilation) to the Nintendo 3DS using AI agents. Orchestrated by a single long-lived Claude session directing 147 subagents, the project achieved 60fps stereoscopic gameplay on New 3DS in 24 days, later expanding to Old 3DS support and native rendering. The effort consumed ~3.17 billion tokens with an API cost of nearly $5,000, demonstrating the feasibility and economics of large-scale autonomous coding agents for complex legacy system migrations.