Intermediate · Systems

Persistent Background Agents

Visual diagram · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. Long-running autonomous agents — such as Meta's Muse Code and Replit's Agent 4 — that build, debug, and ship software over days from a single natural-language instruction.

Technical Definition

Long-running autonomous agents — such as Meta's Muse Code and Replit's Agent 4 — that build, debug, and ship software over days from a single natural-language instruction.

How it works

Persistent background agents break the chat-session model of AI assistance. Instead of answering in one sitting, they accept a goal, decompose it into a plan, and keep working asynchronously — writing code, running tests, hitting errors, recovering, and reporting back hours or days later. The hard engineering problems are state durability (the agent must survive restarts and context-window limits), supervision (humans need checkpoints, diffs, and rollback rather than a firehose of actions), and cost control (an unconstrained agent can burn unbounded compute). Evaluation shifts from answer quality to outcome quality: did the shipped artifact actually work?

Visual Explanation (cycle)

Goal → Plan → Act (code/research) → Observe results → Recover from errors → Checkpoint to human → repeat until done

Related Concepts

• Agent-to-Agent (A2A) Protocol — A standardized framework, generally released in September 2026, that lets separate AI agents securely communicate, delegate tasks, and collaborate across systems.
• Outcome-Based Evaluation — Judging AI systems by the quality of real-world outcomes they produce rather than the surface quality of individual answers.
• Agent Memory — Information an agent retains beyond the current context window — episodic history, learned facts, preferences, and task state.