Coding Assistants: Under an Agentic Regime, a New Landscape

With pricier raw materials, pricier products.

In the summer of 2025, the phenomenon surfaced at Cursor. The “raw material” in question: Anthropic’s LLMs.

As their costs rose, Cursor ultimately shifted its individual plans to a token-based model, moving away from the model focused on requests as untenable.

A few weeks later, Gartner mentioned the episode in the Magic Quadrant overview dedicated to coding assistants. Anthropic, it noted, represented a threat amplified by having positioned itself on this market with Claude Code. The analyst cited this as emblematic of a broader trend: LLM providers increasingly developing their own coding assistants, with a notable advantage in that they can jointly optimize both their models and their agentic harnesses. The competition intensified as publishers responded by building their own models. Cursor serves as an example, leveraging SpaceX’s compute resources for this purpose.

From “assistants” to “agents”: GitLab exits the Magic Quadrant

Last year, the Magic Quadrant focused on “coding assistants.” Gartner had incorporated the agentic aspect into its assessment, but not as a non-negotiable criterion (multiline auto-completion with natural-language suggestions, generation of unit tests and documentation, awareness of internal code-repository context, etc.).

This year, the focus shifts to “coding agents.” The evaluation criteria reflect this shift. Broadly speaking, the following were mandatory:

• Autonomous execution of tasks from natural-language instructions
• Iterative verification and self-correction (build, test, validate, debug, refactor)
• Integration with development tools and environments
• Automatic context management (identification, selection, reuse…)
• Support for MCP
• Human oversight, traceability and audit mechanisms
• Usage analytics
• User and access management, code exclusion controls
• Guarantees against training client code and documentation

Multi-agent orchestration was optional. As, among others:

• Creation of custom sub-agents
• Workflow triggers based on events
• Management of skills (instruction models)
• Unified context orchestration (structured knowledge base construction)
• Spec-driven workflows
• Specialist agents for code modernization and translation
• Code-review assistance
• Cost management
• Multiple deployment models

Of the 14 vendors ranked last year, 6 are no longer listed this year: Augment Code, Harness, IBM, Qodo, Tencent Cloud… and GitLab, which Gartner had positioned among the “leaders.” Not without noting that GitLab was not ahead on several components, including chat, analytic dashboards and self-hosting options.

Amazon, Cognition and Google are no longer “leaders”

The Magic Quadrant is structured around two axes. One, “execution,” reflects the ability to effectively meet demand. The other, “vision,” captures the strategies (commercial, marketing, sectoral, geographic…).

Three “leaders” from last year slipped to “challengers” due to a retreat in vision: Amazon, Cognition and Google.

Amazon, Gartner contends, offers a product that is less differentiated than competing offerings, particularly for those seeking asynchronous agent functionality. At the time of evaluation (closing March 2, 2026), parallel execution of agents was not available. Pricing and packaging remained in flux (pooling and mixed-user populations, in particular). As for Kiro, its spec-driven approach does not always align with needs and working methods.

At Cognition, the value proposition is harder to scale due to the substantial integration and customization work required. Another risk: the billing unit (ACU, Agent Compute Units) remains abstract. Gartner questions Cognition’s overall viability given its low self-service conversion and adoption rates.

At Google, foundational models and agent harnesses still lack cohesion. The array of agentic coding tools overlaps in functionality. And the sector-focused strategy backing the partner network lacks clarity.

Anthropic, Cursor and OpenAI join GitHub among the “leaders”

Alongside the five vendors that dropped out, Gartner records four entrants: Anthropic, Atlassian, BytePlus and OpenAI.

The execution axis shows the following developments:

On the “vision” axis:

Anthropic: operational maturity to be validated after the early-2026 incidents

Anthropic has managed to translate its momentum in foundation-model tooling into one of the most popular coding products, aligned with demand, Gartner notes. It has tightly integrated these foundations with its agent harness Claude Code, thus gaining a structural edge in performance and user experience. The CLI-first strategy, paired with IDE extensions, has also allowed it to reach developers where they are located.

In early 2026, Anthropic experienced service disruptions and release issues. Its post-incident communications were uneven. Against this backdrop, Gartner warns about the need to validate operational maturity for critical workflows. The analyst also highlights vertical integration: users remain bounded to Claude models hosted by Anthropic.

Cursor could advance on enterprise support

Cursor stands out for the depth of its product—from parallel and asynchronous execution to expanding capabilities for testing and code review. Gartner notes the ability to leverage state-of-the-art third-party models, combined with Cursor’s growing investments in its own models. It also points to a meaningful enterprise market share, reinforced by a robust commercial strategy (expanding sales teams, partner networks and geographic reach).

In a market where hyperscalers and large providers (OpenAI, Anthropic) coexist, Cursor remains exposed to margin pressure. Its enterprise support is still evolving. At present, its partner network appears more as an indirect sales channel than a delivery lever. And the six-hour target for first response time remains higher than what other “leaders” commit to.

GitHub: losing mindshare among developers

IDE, CLI, web interface, CI, security… GitHub benefits from being a platform that can operationalize its own agents as well as those from third parties. Gartner also praises its effective integration of agent tooling into existing workflows, and notes the breadth of governance and compliance controls, while reminding that GitHub benefits from the backing of its parent, Microsoft.

GitHub’s leadership in coding agents is not as clear as its dominance in coding assistants. This is especially true for event-driven workflows, asynchronous execution and orchestration of preconfigured agents. Moreover, developers’ brand awareness has softened. And the governance-control focus may not resonate with those seeking primarily functional capabilities.

Ongoing legal and regulatory uncertainties for OpenAI

Like Anthropic, OpenAI gains a structural edge by aligning the capabilities of its models with its Codex harness. It also stands out for governance, including OS-level agent isolation, approval gates, RBAC and workspace-level audits. Its enterprise-focused go-to-market is bolstered by flexible deployment options and a large ecosystem of connectors and partners.

With Codex optimized for GPT models, performance with third-party models will need to be validated. It’s also worth remembering that OpenAI entered this market later than many competitors—and one should verify whether the desired extensions and integrations are sufficiently mature. Legal and regulatory uncertainties surrounding OpenAI—copyright lawsuits, privacy litigations and antitrust inquiries—should also be kept in view.