“What are the components of agent harness and how do they contribute to the harness?”
A synthesis of 10 component groups drawn from live web sources (2026-09-02): a source-code study of 11 coding-agent systems cross-checked against official OpenAI, Anthropic, LangGraph, Google ADK, and Microsoft Agent Framework materials. Eight groups form the core runtime and control surfaces; observability and evaluation close the production feedback loop.
| # | Layer | Component & contribution | Terminology in sources | Pages | Source |
|---|---|---|---|---|---|
| 1 | Foundation | Model integration & instructions. Connects the harness to one or more model providers, assembles instructions and prompts, and handles routing, caching, structured responses, and model-specific protocols. | Augmented LLM, LLM integration, model clients | 3 | arxiv.org |
| 2 | Runtime | Agent loop & runner. Runs the perceive-reason-act cycle: call the model, interpret its decision, execute actions, feed results back, and enforce stop, timeout, cancellation, retry, and failure-recovery conditions. | Agent loop, Event Loop, Runner | 3 | arxiv.org |
| 3 | Capabilities | Tools, actions & environment. Defines what the agent can do beyond text—such as search, code execution, file editing, database/API calls, and interaction with external systems—and returns action results to the loop. | MCP clients, Tools & actions | 2 | arxiv.org |
| 4 | State | Memory, context & retrieval. Builds the model context for each turn, manages the finite context window, retrieves relevant knowledge, and preserves short-term conversation state or long-term knowledge across sessions. | Comprehensive memory, Memory & context, agent session, context providers | 3 | arxiv.org |
| 5 | State | Persistence & durability. Checkpoints execution and state so long-running work can survive failures, pause for external events or people, and resume without starting over. | DBOS, Dapr, Persistence, Restate, Resume Agents, Temporal | 3 | docs.langchain.com |
| 6 | Control | Safety, permissions & human oversight. Constrains risk by validating inputs, outputs, and tool calls; applying least-privilege permissions and sandboxes; setting budgets; and escalating sensitive actions for human approval. | Guardrails, Human-in-the-loop, Input guardrails, Output guardrails, Safety & permissions, Tool guardrails | 4 | arxiv.org |
| 7 | Coordination | Workflow & multi-agent orchestration. Decomposes work, routes tasks, coordinates parallel or sequential steps, delegates to specialized agents, aggregates results, and manages handoffs. | Agents as tools / Handoffs, Orchestration, Orchestrator-workers, Parallelization, Routing, Workflows | 4 | arxiv.org |
| 8 | Integration | Extensibility & integration surface. Lets developers add providers, tools, skills, hooks, middleware, plugins, and standard protocol connections without redesigning the core runtime. | Extensibility, MCP clients, middleware | 2 | arxiv.org |
| 9 | Operations | Observability & tracing. Records end-to-end runs and timed spans across model calls, tools, guardrails, and handoffs so teams can debug behavior, audit decisions, and monitor latency, cost, and failures. | Spans, Traces, traces | 2 | developers.openai.com |
| 10 | Quality | Evaluation & feedback. Uses graders, datasets, replayable eval runs, and feedback loops to measure task success, detect regressions and failure modes, and compare changes before deployment. | Evaluator-optimizer, datasets, eval runs, graders | 2 | developers.openai.com |
Method: 10 normalized component groups synthesized from live web research on 2026-09-02, cross-checked against a source-code study of 11 coding-agent systems and official OpenAI, Anthropic, LangGraph, Google ADK, and Microsoft Agent Framework documentation. "Pages" counts independent web pages supporting each group (2–4 per row). Frameworks use differing labels, so the groups are a synthesis, not a claim of identical terminology; full supporting URL lists were trimmed to one primary link per row for space.