Five-axis classification of the agent landscape as of February 2026. The "coding copilot" category is dissolving; what replaces it is broader, more autonomous, and aimed at all knowledge work.
The landscape divides into six domains. These are not mutually exclusive -- several products span multiple domains -- but they represent distinct problem spaces with different user personas and evaluation criteria. Click any product tag to see details and highlight across all views.
Inline code suggestions, completions, and chat-based code generation within an IDE. Reactive. Tightly coupled to editor context. Commoditising.
Autonomous multi-step code generation, testing, debugging, deployment. Plans, executes, verifies, iterates. Minutes to hours. Current battleground.
Multi-step non-coding tasks: file organisation, document creation, data analysis, expense processing. The category Cowork created. This is where the "SaaS displacement" thesis lives.
Navigate web browsers or desktop applications as a human would -- clicking, typing, scrolling, reading screens. Authentication boundary is the critical architectural decision.
Search, synthesise, and produce structured reports from multiple sources. Evaluated on citation quality, source diversity, and synthesis depth.
AI-powered vulnerability scanning positioned against traditional SAST/DAST. New competitive axis -- CrowdStrike -8%, Cloudflare -8.1% on Claude Code Security launch.
Adapted from the academic literature (Sapkota et al. 2025) but calibrated to observed product behaviour. Higher autonomy = higher capability and higher risk.
The most significant structural development: the dissolution of the boundary between "coding tool" and "knowledge work tool." The coding copilot was always a trojan horse for a general-purpose agent platform.
Click events to expand details
Anthropic now has the broadest deployment surface of any agent vendor: terminal (Claude Code), desktop (Cowork), browser (Chrome), Office (Excel, PowerPoint), Slack, web (claude.ai), mobile (Claude app), and IDE (Xcode). The underlying agent architecture is domain-agnostic; the surfaces are presentation layers. No other vendor spans this range.
Anthropic's strategy is horizontal: one agent, many surfaces. Microsoft's strategy is vertical: one agent per application. Google's strategy is embedded: AI native to each Workspace app.
"Coding copilot" and "productivity suite AI" are converging. Procurement should treat them as a single landscape.
The underlying agent loop (plan, execute, verify, iterate) is identical whether refactoring code or reconciling invoices.
Developer-adopted tools (Claude Code, Cursor) are expanding into non-developer workflows. Coding tool contracts will grow into enterprise-wide negotiations.
How the major vendors are positioning. Each represents a distinct bet on where agent value concentrates and how it is captured.
One agent architecture (Claude Agent SDK), many surfaces. Extend from developer base into general knowledge work via Cowork. MCP as ecosystem play.
Rapid model iteration (4 Codex variants in 6 weeks). Cerebras partnership for 1000+ tok/s. Codex as general computer operation agent.
AI embedded in every M365 application. Copilot Studio for custom agents. Graph for data grounding. Now offers Claude models within Copilot.
Gemini in Workspace at aggressive pricing ($14/user/mo). CLI open-sourced. Vertex AI for custom agents. Gems and Agentspace for no-code agents.
Model-agnostic, no vendor lock-in. OpenCode (95K stars), Ollama (NVIDIA partnership), LM Studio (headless daemon). 42%+ of developers running LLMs locally.
Click a vendor to highlight their products across all tabs. Press ESC to clear.
Major products mapped against six capability domains. Filled circles indicate primary capability; half-filled indicates partial or emerging.
| Product | Vendor | Code Gen | Agentic Code | Knowledge Work | Browser | Research | Security |
|---|---|---|---|---|---|---|---|
| Claude Code | Anthropic | ||||||
| Cowork | Anthropic | ||||||
| Claude in Chrome | Anthropic | ||||||
| Claude in Excel | Anthropic | ||||||
| OpenAI Codex | OpenAI | ||||||
| OpenAI Operator | OpenAI | ||||||
| Cursor | Anysphere | ||||||
| GitHub Copilot | Microsoft | ||||||
| M365 Copilot | Microsoft | ||||||
| Gemini Workspace | |||||||
| Gemini Code Assist | |||||||
| Jules | |||||||
| Manus | Meta (acq.) | ||||||
| Devin | Cognition AI | ||||||
| OpenCode | Open source | ||||||
| Ollama | Open source | Local inference runtime -- enables all domains via compatible models | |||||
| LM Studio | LM Studio | Local inference runtime -- enables all domains via compatible models | |||||
How the taxonomy maps to the Peerlabs Four Axes methodology for enterprise technology assessment.
Model capabilities converging at code generation level; differentiating on agentic reasoning, multi-step execution, and context window. Autonomy level is the key functional axis -- Level 2-3 agents are where value and risk concentrate.
Enterprise adoption signals strong (NVIDIA, Dropbox, Cisco, Salesforce). But Cowork's enterprise readiness gaps (no audit logs, local-only plugins) suggest knowledge work agents are 6-12 months from enterprise-grade. M365 Copilot and Gemini are enterprise-ready but less capable.
MCP achieving protocol-level ubiquity. ACP and AGENTS.md emerging as complementary standards. OpenAI's App Server is the notable holdout. Integration architecture decisions in the next 12 months will shape vendor lock-in for 5+ years.
The "vibe coding" debate extends to "vibe working" -- Cowork enables non-developers to execute complex multi-step workflows. Workforce implications are being priced into markets. Security and governance lag capability by 6-12 months.