Agentic Harnesses Market Landscape, July 2025 to July 2026

Narrative

Practitioner evidence separates supervised coding harnesses from background software-engineering agents more clearly than it measures vendor market share. Thoughtworks’ April 2025 Radar placed Cursor, Cline and Windsurf among supervised IDE-centred agents, while Birgitta Böckeler distinguished these interactive tools from autonomous background agents such as Codex, Jules and Devin that work in isolated environments and commonly return pull requests. GitHub, OpenAI and Google subsequently converged on the same operational pattern: issue or prompt delegation, a managed or sandboxed execution environment, tests and a reviewable pull request. (thoughtworks.com)

Adoption evidence is stronger for broad AI assistance than for any one harness. DORA’s 2025 report characterises AI as an amplifier of pre-existing organisational capabilities, and its 2026 follow-up reports 90% workplace use among technology professionals while warning that higher adoption can raise both throughput and delivery instability. Stack Overflow’s 2025 survey similarly finds that agents remain non-mainstream, with 52% either not using them or sticking to simpler tools, but 84% of software-developer agent users employ them for software development; its out-of-the-box tool responses put ChatGPT at 82% and GitHub Copilot at 68%, a useful but non-equivalent proxy for coding-harness reach. (dora.dev)

The durable differentiation is increasingly in the harness rather than the chat interface: repository and environment context, permissioning, test and CI feedback, state persistence, traceability, cost control and approval loops. OpenAI’s Codex engineering accounts describe an explicit agent loop and a shift towards engineers designing environments and feedback loops, while Anthropic exposes related context-management, permission and checkpointing primitives through its Agent SDK. GitHub’s CI/CD material and Thoughtworks’ 2026 Radar make the corresponding governance point: useful agents are permission-hungry, so constrained containers, least privilege, audit trails and human gates are design requirements rather than add-ons. (openai.com)

For broader workflow agents, the production architecture centres on explicit state and control flow, not merely a collection of role prompts. LangGraph argues for cyclic, deterministic graph execution with durable checkpoints, while Microsoft’s successor framework to AutoGen shifts multi-agent coordination towards typed, graph-based workflows and makes the agent-versus-workflow choice explicit. The available sources support planner-worker, sub-agent and human-review patterns, but not a general claim that agent swarms outperform carefully bounded single-agent or deterministic workflow designs; cost, hidden state, prompt injection and verification remain recurrent constraints. (langchain.com)