Agent Edge | May 28, 2026
🧠 Founder runs a company with 89 AI agents across 22 departments
@JaredSanborn | Reddit
🔗 https://www.reddit.com/r/AI_Agents/comments/1tp5zw9/i_run_a_company_with_89_ai_agents_across_22/
A founder running a real company on 89 AI agents posted detailed field notes on what actually works. The biggest finding: delegation is the bottleneck, not intelligence. Agents are smart enough, but knowing which agent to invoke and how to coordinate outputs is the hard part. The solution was a dedicated “conductor” agent that does nothing but orchestration. Department structure with manager agents beat flat coordination by a wide margin. The founder describes agents needing roughly 100 invocations with memory before becoming genuinely useful.
📌 Why it matters: The field has spent two years chasing smarter models while this founder proves the real unlock is coordination architecture. A conductor agent is an organizational pattern, not a model improvement. The 100-invocation threshold is also a practical signal: expect cold-start friction when deploying new agents and budget for that ramp. Multi-agent systems fail on structure, not capability.
🤖 Agent angle: Before adding another model to your stack, audit your coordination layer. Does each agent know who to hand off to? Is there a routing mechanism or are you hardcoding chains? Build a conductor agent as a separate service with its own memory before you scale past five agents. And plan for a 100-invocation onboarding period before judging any new agent’s performance.
🌐 The internet is being rebuilt for machines
Rebecca Bellan | TechCrunch
🔗 https://techcrunch.com/2026/05/28/the-internet-is-being-rebuilt-for-machines/
AWS redesigned OpenSearch Serverless specifically for agentic workloads, decoupling compute from storage so infrastructure scales from zero to burst in seconds and back to zero when idle. Cloudflare reports bots already account for 31% of HTTP traffic, with a senior PM predicting non-human traffic will exceed human traffic in the first half of 2027. Google I/O showed users delegating tasks like travel booking and research directly to AI agents. AWS’s Tia White described the pattern: agents spike without warning and go idle without notice, making traditional pay-for-provisioned-capacity models wasteful.
📌 Why it matters: The infrastructure layer is pivoting to agent-shaped economics. Zero-to-burst scaling is table stakes for any service that wants to serve agent traffic, because agents do not follow human traffic patterns. Cloudflare’s 31% bot traffic number is already straining legacy rate-limiting and authentication models. Every team deploying agents should expect their infrastructure costs to decouple from traditional web traffic baselines.
🤖 Agent angle: Rethink your service architecture around spiky, unpredictable demand. If your database or search backend provisions capacity upfront, you will overpay or hit cold-start latency when your agents scale. Evaluate OpenSearch Serverless and similar zero-to-scale services for any backend your agents call directly. Also monitor your own agent traffic: if your agents call your APIs, those calls are bot traffic and need separate rate limiting from user traffic.
🚀 Claude Code ships dynamic workflows with parallel subagent fleets
@ClaudeDevs | X/Twitter
🔗 https://x.com/ClaudeDevs/status/2060044853279617150
Claude Code launched dynamic workflows as a research preview. The agent writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to tackle tasks too big for a single agent loop. Use cases include service-wide bug hunts, large code migrations, and stress-testing architectural designs. Just include the word “workflow” in a prompt to start using it. The post has 2.1 million views and 8,300 likes.
📌 Why it matters: This is the first major shipping product where agents write their own orchestration code at runtime. The “just say workflow” entrypoint is a UX breakthrough: it hides an enormous amount of complexity behind a single keyword. Parallel subagent fleets are a new primitive that changes the economics of large-scale refactoring and testing. If Claude Code can dynamically split and delegate work correctly, it sets a new bar for what a coding agent can accomplish in a single session.
🤖 Agent angle: Try the workflow feature on your largest codebase task this week. Start with something like “find every unhandled error path in the auth module and fix them, using workflow.” Observe how the agent decomposes the task and how subagents communicate results back. The orchestration script it generates is worth studying as a reference for your own multi-agent designs.
🛠️ PilotDeck goes open source: task-oriented agent OS from Tsinghua and OpenBMB
OpenBMB | GitHub
🔗 https://github.com/OpenBMB/PilotDeck
PilotDeck, an open-source agent operating system built by Tsinghua University’s THUNLP, OpenBMB, and ModelBest, hit 1,700 GitHub stars in its first days after open-sourcing on May 28. Designed around WorkSpaces that completely isolate files, memory, and skills per project. Key features include traceable white-box memory that lets you pinpoint and edit individual memory entries, smart model routing that assigns different difficulty tasks to different models, and an always-on mode. Natively supports MCP.
📌 Why it matters: Agent operating systems are becoming a category, and PilotDeck’s WorkSpace isolation model solves a real pain for teams running multiple agents in the same deployment. Contamination between agent memories is a quiet but persistent failure mode that PilotDeck addresses head-on. White-box memory is a differentiator: most agent memory systems are black boxes you cannot debug. Smart model routing also matters economically, letting you pay for expensive reasoning models only on hard tasks.
🤖 Agent angle: Clone the repo and evaluate WorkSpace isolation for your multi-project agent setup. If you have ever seen Agent A’s state leak into Agent B’s context, PilotDeck’s approach is worth adopting. The white-box memory feature alone justifies a test drive, because debugging opaque agent memory is one of the hardest problems in production. Run the always-on mode against a simple recurring task to see how it handles sustained operation.
💡 Trajectory raises $15M to build continual learning for deployed agents
@ZeffMax | X/Twitter
🔗 https://x.com/ZeffMax/status/2059639290280497228
A group of AI researchers from Google DeepMind, Apple, Microsoft Research, and OpenAI launched Trajectory, a startup building a continual learning platform for companies deploying AI agents. Raised $15M seed round led by Sarah Guo’s Conviction, with participation from Jeff Dean and Fei-Fei Li. The premise: most deployed agents never improve from user corrections because feedback never flows back into training. Trajectory’s platform post-trains models on real human corrections captured in production.
📌 Why it matters: This is the first serious attempt to close the feedback loop that every production agent system currently leaks. Right now, when a user corrects an agent’s output, that correction vanishes into a log file or, at best, a short-term prompt update. Trajectory’s bet is that the company that captures and trains on production corrections will build a compounding advantage over competitors whose agents repeat the same mistakes. The founding team’s pedigree signals that the frontier labs see this as a solvable problem, not a research moonshot.
🤖 Agent angle: If you run agents in production today, start capturing every user correction in a structured format now, even without Trajectory’s platform. Tag each correction with the original prompt, the agent’s output, the user’s fix, and the task type. This data is your moat. When post-training pipelines become plug-and-play (which Trajectory and others will make inevitable), the teams with clean correction datasets will pull ahead of everyone else.