Manu's Vault

AI-Coding-Agent-Architecture

3 min readStatus:Seedling

An AI coding agent is not merely a language model — it is a compound system combining a reasoning model with surrounding infrastructure that enables autonomous action in a software environment.

The foundational equation is:

coding agent = AI model(s) + harness

The model provides language understanding, reasoning, and code generation. The harness provides everything the model needs to act: tools, configuration, memory, verification, and environmental context.

Core Architectural Components

Modern AI agent architecture comprises four subsystems:

  • Perception: Captures and processes environmental inputs — file contents, terminal output, test results, web data — transforming them into representations the model can reason over
  • Reasoning/Planning: Decomposes complex tasks into subtasks, generates solution candidates, and reflects on outcomes; techniques include Chain-of-Thought, Tree-of-Thought, and ReAct (Reason + Act)
  • Memory: Retains knowledge across contexts — short-term memory lives in the context window; long-term memory is stored in files, databases, or vector indices and retrieved as needed
  • Execution/Action: Translates model decisions into concrete environment changes — running shell commands, editing files, calling APIs, or spawning sub-processes

What Distinguishes Agents from Chatbots

The defining capability is tool use in a loop. As Mitchell Hashimoto articulates: “An agent is the industry-adopted term for an LLM that can chat and invoke external behavior in a loop.” Minimum capabilities include reading files, executing programs, and making HTTP requests.

Chatbots respond; agents act. When given a way to verify their own work, agents can detect errors and self-correct without human prompting. This feedback loop — generate, execute, observe, revise — is absent from conversational interfaces.

Historical Lineage

Pre-LLM software agents used deliberative architectures like BDI (Belief-Desire-Intention), where:

  • Beliefs represent the agent’s world-state knowledge
  • Desires represent goals the agent pursues
  • Intentions represent committed plans of action

BDI agents could reason and plan, but required handcrafted knowledge representations and explicit programming of every capability. LLMs collapsed this complexity: the model itself handles natural language understanding, knowledge retrieval, and plan generation — all previously separate subsystems. The harness pattern emerged to give LLMs grounding in real environments that BDI systems previously hardcoded.

Why Architecture Matters

Understanding the model + harness decomposition is essential for Harness-Engineering. The model is largely fixed; the harness is what practitioners engineer. Context-Engineering — how information is structured and delivered to the model — operates as the primary lever within this architecture.

Sources

  • Horthy, Dex (2026). “Skill Issue: Harness Engineering for Coding Agents.” HumanLayer Blog. Retrieved from https://www.humanlayer.dev/blog/skill-issue-harness-engineering-for-coding-agents

    • Source of the core coding agent = AI model(s) + harness equation and harness component taxonomy

  • Hashimoto, Mitchell (2026). “My AI Adoption Journey.” mitchellh.com. Retrieved from https://mitchellh.com/writing/my-ai-adoption-journey

    • Practitioner definition of agent as “LLM that can chat and invoke external behavior in a loop”; tool-use as the key differentiator from chatbots

  • Weng, Lilian (2023). “LLM Powered Autonomous Agents.” Lil’Log. Retrieved from https://lilianweng.github.io/posts/2023-06-23-agent/

    • Widely cited framework establishing the planning, memory, and tool-use triad as the canonical LLM agent architecture

  • Zhang, Yizhang, et al. (2025). “Fundamentals of Building Autonomous LLM Agents.” arXiv:2510.09244. Retrieved from https://arxiv.org/html/2510.09244v1

    • Academic survey formalizing the four-subsystem model: perception, reasoning, memory, execution

  • Rao, Anand S. and Michael P. Georgeff (1995). “BDI Agents: From Theory to Practice.” Proceedings of the First International Conference on Multi-Agent Systems (ICMAS). AAAI Press. Retrieved from https://cdn.aaai.org/ICMAS/1995/ICMAS95-042.pdf

    • Foundational paper on BDI agent architecture; establishes historical lineage of autonomous agent design prior to LLMs

Note

This note was researched and drafted with AI. How these notes are written →

Graph View

Backlinks