The paradigm is shifting from prompt-driven chatbots to autonomous agent architectures that operate continuously in cloud environments.
Anything (15-person team, 1M+ users) is running 150+ specialized agents with persistent state, memory, and dedicated compute instances. These aren't wrapper scripts—they're narrow-scope agents with individual tool access, Slack integration, and async execution models.
Key architectural shift: Instead of monolithic multi-tool agents, they're deploying role-specific agents (Emma for support, Grace for bug triage, Canary for prod monitoring). Each agent maps to a traditional job function with scoped permissions and context windows.
The interface layer is evolving beyond web chat: CLI integration for dev workflows (works alongside Cursor/Claude), iMessage endpoints for mobile access, and Slack profiles with real avatars. Agents persist across sessions and maintain conversation history without manual context injection.
Multi-agent coordination is still rough—group chats show agents can collaborate but lack decisive convergence mechanisms. They're overly deferential and don't always reach final states without human intervention.
The compute model matters: Each agent runs on its own cloud instance with browser automation, file system access, and API credentials. This is fundamentally different from stateless LLM calls.
Cost implications: Narrow agents on cheaper models vs. one powerful agent on frontier models. The economics favor specialized deployment as o1/o3 pricing scales.
Bottom line: We're moving from "what can a model do per prompt" to "what can an agent accomplish autonomously over time." The productization of agentic systems is happening faster than the research community expected.
Anything (15-person team, 1M+ users) is running 150+ specialized agents with persistent state, memory, and dedicated compute instances. These aren't wrapper scripts—they're narrow-scope agents with individual tool access, Slack integration, and async execution models.
Key architectural shift: Instead of monolithic multi-tool agents, they're deploying role-specific agents (Emma for support, Grace for bug triage, Canary for prod monitoring). Each agent maps to a traditional job function with scoped permissions and context windows.
The interface layer is evolving beyond web chat: CLI integration for dev workflows (works alongside Cursor/Claude), iMessage endpoints for mobile access, and Slack profiles with real avatars. Agents persist across sessions and maintain conversation history without manual context injection.
Multi-agent coordination is still rough—group chats show agents can collaborate but lack decisive convergence mechanisms. They're overly deferential and don't always reach final states without human intervention.
The compute model matters: Each agent runs on its own cloud instance with browser automation, file system access, and API credentials. This is fundamentally different from stateless LLM calls.
Cost implications: Narrow agents on cheaper models vs. one powerful agent on frontier models. The economics favor specialized deployment as o1/o3 pricing scales.
Bottom line: We're moving from "what can a model do per prompt" to "what can an agent accomplish autonomously over time." The productization of agentic systems is happening faster than the research community expected.