Resources for building an agentic harness.

Coding-agent IDEs and CLIs

The tools that put an agent inside your editor or terminal. Pick one for daily work; nothing on this page assumes a particular choice. The harness pattern is portable across all of them, and the build-your-own playbook explains how.

• Cursor

  What
  IDE fork of VS Code with a first-class agent panel, plan mode, and .cursor/rules for persistent per-project guidance.

  When you would pick it
  You want an editor-native agent with strong keyboard workflow and shared team rules under version control.

• Cursor Rules

  What
  The rules system: project rules in .cursor/rules/*.mdc, user rules, team rules, and AGENTS.md fallback, with glob-scoped activation.

  When you would pick it
  You already write Cursor and want to encode standards the agent applies every session without reminding it.

• Claude Code

  What
  Anthropic's terminal, IDE, desktop, and web coding agent. Reads the codebase, edits files, runs commands, and integrates with git.

  When you would pick it
  You want an agent that operates across surfaces (CLI, VS Code, JetBrains, browser) with the same session, hooks, and skills.

• Claude Agent SDK

  What
  Python and TypeScript library for building custom agents on top of Claude Code's agent loop, tools, hooks, subagents, and permissions.

  When you would pick it
  You are past interactive use and want to embed the agent loop in your own process, service, or CI.

• GitHub Copilot

  What
  GitHub's in-editor assistant plus Copilot Chat, code review, and custom agents that run in the browser and IDE.

  When you would pick it
  You live inside GitHub Enterprise, need SSO / audit / policy controls, and want an agent wired directly into PRs and Actions.

• OpenAI Codex CLI

  What
  Open-source terminal coding agent from OpenAI. Runs locally with your ChatGPT account, with IDE integrations for VS Code, Cursor, and Windsurf.

  When you would pick it
  You are on the OpenAI stack and want a CLI-first agent without leaving your terminal.

• Aider

  What
  Open-source terminal pair programmer. Model-agnostic, git-aware, auto-commits every change with a descriptive message.

  When you would pick it
  You want a minimal, script-friendly agent that treats git as the ledger and works with any capable model.

• OpenHands

  What
  Open-source self-hosted agent runtime (formerly OpenDevin). Runs Claude Code, Codex, Gemini, or its own agents against local, remote, or cloud backends.

  When you would pick it
  You need a multi-agent control plane you can host yourself, wire into Slack, GitHub, or Linear, and trigger on a schedule.

Building agents: concepts and walkthroughs

The papers and posts that explain why agent loops are shaped the way they are. Read at least the first two before you reach for a framework.

• Building effective agents (Anthropic)

  What
  The reference primer on agent patterns: augmented LLMs, prompt chaining, routing, parallelization, orchestrator-worker, evaluator-optimizer, and autonomous loops.

  When you would pick it
  You are about to reach for a framework. Read this first. Most of what you need is a small composition of the patterns here.

• A harness for every task (Claude)

  What
  How the Claude Code team frames dynamic multi-agent workflows: fan-out-and-synthesize, adversarial verification, tournaments, loop-until-done, generate-and-filter.

  When you would pick it
  You already have a working single-agent flow and are asking whether it is time to orchestrate multiple.

• ReAct: Reasoning and Acting in LMs (arXiv)

  What
  The paper that popularized interleaving reasoning traces with tool-use actions. The foundation nearly every agent loop still reflects.

  When you would pick it
  You want to understand why agent loops are shaped the way they are before you go designing a new one.

• OpenAI Agents SDK

  What
  OpenAI's lightweight agent framework: agents, tools, sandboxes, realtime and voice variants, guardrails. Small surface area, fast to read.

  When you would pick it
  You want a minimal building block from the OpenAI side, without pulling in a heavier orchestration graph.

• LangGraph (LangChain)

  What
  State + nodes + edges. A message-passing graph model for building durable, controllable multi-step agent workflows with checkpointing and streaming.

  When you would pick it
  Your agent has enough branching, retries, and human-in-the-loop pauses that you want a real graph runtime holding it together.

• LangChain agents

  What
  The classic model-calls-tools-in-a-loop building block, now recommended for smaller flows before you graduate to LangGraph.

  When you would pick it
  You want the shortest path from a working prompt to a tool-using agent inside an existing LangChain project.

Harness patterns and durable execution

An agent that runs for five minutes is a script. An agent that runs for five hours is a workflow. These runtimes handle the second case: journaled steps, deterministic replay, human-in-the-loop pauses, structured retries. Any of them can act as the outer harness around an LLM loop.

• Temporal for AI

  What
  Durable execution runtime pitched as the orchestrator for AI applications. Long-running sessions survive crashes, retries, and rate limits by replaying journaled steps.

  When you would pick it
  Your agent needs to run for hours or days, pause for human approval, and resume after any kind of failure without losing state.

• Temporal AI cookbook

  What
  Concrete recipe: OpenAI Agents SDK activities wrapped as Temporal tools, so your agent loop inherits deterministic replay and structured retries.

  When you would pick it
  You already know Temporal and want to plug an agent SDK into it without inventing your own activity boundaries.

• Restate

  What
  Single-binary durable execution and virtual-object runtime. Ships as its own agent runtime with memory, human-in-the-loop pauses, and long-running tasks.

  When you would pick it
  You want durable execution without operating a full workflow cluster, and you like the actor / virtual-object model for agent state.

• DBOS

  What
  Open-source durable execution library that hooks into Postgres. Native integrations with Pydantic AI, LlamaIndex, and the OpenAI Agents SDK.

  When you would pick it
  Your stack is already Postgres-centric and you want durable workflows without a separate orchestration service.

Quick answers

The five questions readers land on this page asking, answered in one paragraph each.

What is an agentic harness?

The scaffolding around an LLM: rules, hooks, gates, memory, permissions, and multi-agent orchestration. The LLM is the engine; the harness is the vehicle. A harness turns a capable model into a reliable teammate by encoding your standards and enforcing them mechanically.

Which AI coding IDE should I pick?

Pick Cursor for an editor-native agent with .cursor/rules for team standards. Pick Claude Code for a multi-surface agent (CLI, IDE, desktop, web) with hooks, skills, and subagents. Pick GitHub Copilot if you live inside GitHub Enterprise. Pick OpenAI Codex CLI or Aider for terminal-first workflows. Pick OpenHands to self-host a multi-agent runtime.

What is durable execution and why does it matter for agents?

A runtime pattern in which every step of a workflow is journaled so the workflow can resume exactly where it left off after a crash, restart, or rate limit. For agents this matters because long-running loops call flaky tools, hit quota, and pause for human approval. Temporal, Restate, and DBOS are three production-grade options.

Do I need a framework to build an agent?

No. Anthropic's Building Effective Agents post explicitly recommends against reaching for a framework by default. Start with a simple loop: model calls tools, tools return results, loop continues until the model signals completion. Add a framework only when you have a specific gap it fills (durable execution, graph-shaped control flow, cross-agent handoff).

What is llms.txt?

An emerging convention (analogous to robots.txt) that publishes a plain-text summary of a site aimed at LLM crawlers and AI answer engines. It lives at /llms.txt and points at the pages worth indexing. paw publishes one at getpaw.dev/llms.txt. The convention is documented at llmstxt.org.