Agentic Harnesses Market Landscape, July 2025 to July 2026

Agentic harnesses from July 2025 to July 21, 2026: competitive landscape, market share by use case, model integration, orchestration patterns, and differences between coding harnesses and broader workflow agents, including Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Google Jules, Replit Agent, Devin, Windsurf, OpenCode, Aider, OpenHands, LangGraph, CrewAI, AutoGen, and n8n.

GPT-5.6-sol
financial
frontier
academic
vc
blogs
tech

Synthesised 2026-07-21

Narrative

This lane finds that frontier labs increasingly treat the harness as a first-class product layer rather than a thin chat interface. Anthropic moved Claude Code from terminal-based agent to managed enterprise surface, adding remote MCP connectivity, IDE support, checkpoints, policy controls and usage analytics. OpenAI made an analogous move with Codex across terminal, IDE, cloud, Slack and a desktop application for parallel agents, then exposed much of the underlying pattern through its Agents SDK and AgentKit.

The dominant production pattern is bounded autonomy, not unattended general agency. The relevant releases combine repository or document context, shell and file tools, instructions such as AGENTS.md, sandboxing, approval gates, policy-managed network access, tool connectors, tracing and resumable state. OpenAI's April 2026 SDK release is unusually explicit that durable execution requires separating agent state from disposable compute, while Anthropic's Claude Code materials emphasise remote MCP, subagents and enterprise restrictions.

Coding remains the clearest commercial wedge because repositories, tests, diffs and pull requests supply relatively concrete feedback loops. Yet the frontier-lab evidence also shows expansion into wider knowledge and workflow work: OpenAI reports internal uptake by Legal and Recruiting, Google DeepMind describes multi-agent scientific hypothesis generation, and Anthropic is positioning Claude Code as an execution layer connected to business systems. These are vendor-reported adoption and performance indicators, not independent market-share measures, so they should not be treated as a common market-share series.

Safety and operations have become competitive features of the harness. Codex and Claude Code now package security-review agents, but both vendors stress human review, isolation, approvals and audit trails. METR's February 2026 update supplies a useful external counterweight: agentic tools became common enough among open-source developers to complicate controlled productivity research, which supports a claim of adoption momentum but does not establish that these tools improve developer throughput in all settings.

Sources

ID	Title	Outlet	Date	Significance
t1	Codex is now generally available	OpenAI	2025-10	OpenAI's October 2025 release defines Codex as a unified coding agent across editor, terminal and cloud, and adds Slack delegation, an SDK and administrator controls for managed deployment.
t2	Introducing upgrades to Codex	OpenAI	2025-09	OpenAI describes GPT-5-Codex as a model optimised for interactive pairing and long-running independent coding, with sandboxing, configurable network access and broad deployment surfaces.
t3	Addendum to GPT-5 system card: GPT-5-Codex	OpenAI	2025-09	This safety document records the product-level mitigations for GPT-5-Codex, including agent sandboxing and configurable network access, rather than only marketing claims about coding performance.
t4	Introducing AgentKit	OpenAI	2025-10	OpenAI's AgentKit announcement documents its shift from an SDK to visual workflow construction, guardrails, evaluations and versioning for enterprise agents beyond software engineering.
t5	OpenAI co-founds the Agentic AI Foundation under the Linux Foundation	OpenAI	2025-12	The announcement supplies a dated interoperability signal: OpenAI, Anthropic and others placed AGENTS.md under neutral stewardship and reported adoption across major coding-agent products.
t6	Claude Code and new admin controls for business plans	Anthropic	2025-08	Anthropic details the enterprise harness features that distinguish a managed coding agent from a consumer CLI, including policy settings, MCP configuration controls, spend caps, analytics and a Compliance API.
t7	Enabling Claude Code to work more autonomously	Anthropic	2025-09	Anthropic's September 2025 update adds a native VS Code extension, an updated terminal interface and checkpoints, tying longer autonomous runs to recoverable human supervision.
t8	Anthropic acquires Bun as Claude Code reaches $1B milestone	Anthropic	2025-12	This is Anthropic's strongest public commercial adoption claim for Claude Code, stating that the product reached a $1 billion revenue run rate six months after general availability.
t9	Introducing Labs	Anthropic	2026-01	Anthropic reports Claude Code's rapid commercial growth and MCP's reported monthly-download scale, while situating Claude Code beside Cowork and Skills as experimental agent-product infrastructure.
t10	Claude Code Advanced Patterns: Subagents, MCP, and Scaling to Real Codebases	Anthropic	2026-03	Anthropic's practitioner material provides direct evidence of its recommended orchestration patterns, namely subagents, MCP tool connections and scaling techniques for larger repositories.
t11	Our framework for developing safe and trustworthy agents	Anthropic	2025	Anthropic frames prompt injection, compromised tools and subagents as agent-specific risks, and links MCP directory review to security, safety and compatibility requirements.
t12	How Claude Code is used in practice	Anthropic	2026-06	Anthropic's June 2026 research analyses approximately 400,000 Claude Code sessions and finds a division in which users largely choose objectives while Claude largely determines execution.
t13	Introducing the Codex app	OpenAI	2026-02	OpenAI presents the desktop Codex application as a command centre for multiple parallel agents, using separate project threads and worktrees to avoid repository conflicts.
t14	The next evolution of the Agents SDK	OpenAI	2026-04	OpenAI's April 2026 SDK update is a primary technical account of a model-native harness with configurable memory, MCP, AGENTS.md, filesystem tools, sandbox execution, snapshotting and rehydration.
t15	Running Codex safely at OpenAI	OpenAI	2026-05	This operational account describes the controls that make a coding harness deployable in an organisation: approval policies, sandboxes, network policies, managed configuration and agent-aware telemetry.
t16	Building a safe, effective sandbox to enable Codex on Windows	OpenAI	2026-05	OpenAI's engineering article gives unusually concrete implementation evidence on local-agent isolation, including restricted Windows identities, write limits, firewall rules and process-boundary enforcement.
t17	Codex Security: now in research preview	OpenAI	2026-03	The release shows coding harnesses expanding into application security through repository-specific threat models, sandboxed validation and proposed patches for human review.
t18	How agents are transforming work	OpenAI	2026-06	OpenAI reports internal Codex adoption across technical and non-technical departments, including task-duration evidence, but the figures are company-specific and should not be read as independent market share.
t19	Meet Jules' sharpest critic and most valuable ally	Google Developers Blog	2025-08	Google's Jules update documents critic-augmented generation, an embedded adversarial review loop intended to catch defects and untested assumptions before a coding task completes.
t20	Gemini 3.5 Flash Model Card	Google DeepMind	2026-05	Google DeepMind's model card provides disclosed agentic-tool-use and terminal-coding evaluations, plus distribution through its enterprise agent platform and Google Antigravity.
t21	Co-Scientist: A multi-agent AI partner to accelerate research	Google DeepMind	2026-05	Google DeepMind sets out a concrete multi-agent research pattern: generation and diversity agents followed by reflection, pairwise ranking and simulated debate.
t22	We are Changing our Developer Productivity Experiment Design	METR	2026-02	METR provides independent evidence that uptake of tools such as Claude Code and Codex among open-source developers altered recruitment and experimental conditions, while not claiming a universal productivity gain.
t23	Frontier Risk Report (February to March 2026)	METR	2026-03	METR's external risk report evaluates frontier models on autonomy-relevant software and cyber tasks, supplying a counterweight to lab-reported agent capabilities and safety claims.
t24	Making frontier cybersecurity capabilities available to defenders	Anthropic	2026-02	Anthropic's Claude Code Security preview demonstrates a specialised coding-agent workflow for contextual vulnerability discovery and patch suggestion, with controlled enterprise access.