Agents 2.0: From Shallow Loops to Deep Agents (www.philschmid.de)

The canonical AI agent — a simple while-loop that sends prompts to an LLM, executes tool calls, and repeats — is being supplanted by "Deep Agents" (Agents 2.0), an architectural shift that enables reliable long-horizon, multi-step workflows. Shallow agents work well for short, transactional tasks but fail on complex jobs (hundreds of steps) due to context overflow, loss of the original goal, infinite loops, and poor recovery. Deep Agents address these failure modes by treating planning, state, and execution as first-class system components rather than ephemeral conversation history. Technically, Deep Agents rest on four pillars: explicit planning (maintained to-do plans with statuses and error-handling), hierarchical delegation (an Orchestrator delegates to specialized sub-agents like Researchers, Coders, Writers, each with a clean context), persistent memory (filesystem or vector DBs used as durable state and intermediate-result stores—frameworks like Claude Code and Manus expose read/write access), and extreme context engineering (long, precise prompts, protocol definitions, tool specs, file-naming standards, and human-in-loop formats). The result is modular, auditable, and scalable agent behavior that can tackle tasks spanning hours or days, but it also raises engineering needs around orchestration, tooling, evaluation, and safety as teams standardize protocols and infrastructure.