Intelligence Brief / Q1 2026

Agents at the Gate

The dissolution of the coding copilot category. What began as IDE-embedded code completion has bifurcated into agentic coding and general-purpose knowledge work agents.

Publication February 2026
Coverage Nov 2025 - Feb 2026
Sources 100+
Four Axes F, A, S, P&P

Table of Contents

Start Here
Landscape
Analysis
Decisions
Executive Summary
  • The "coding copilot" category is dissolving. What began as IDE-embedded code completion has bifurcated into developer-centric agentic coding and general-purpose knowledge work agents.
  • Anthropic's progression from Claude Code through Cowork to Chrome, Excel, and PowerPoint integrations demonstrates a single agent architecture expanding from developer tooling into general knowledge work. Bloomberg attributed a $285B software stock selloff to Cowork's launch.
  • Verification and compliance are emerging as a critical gap. Google's Conductor automates post-implementation review. Amazon's Kiro caused two AWS outages when operating autonomously. No vendor has a complete answer.
  • Coding is becoming a shared capability across roles rather than the exclusive domain of software engineers. This is not a 5-year forecast; Cowork already enables non-developers to execute multi-step workflows.
Key Findings

Three Headline Findings

Finding 1
The Escape from the IDE
Anthropic's progression from Claude Code (terminal, mid-2025) through Cowork (desktop, Jan 2026) to Chrome, Excel, and PowerPoint integrations demonstrates a single agent architecture expanding from developer tooling into general knowledge work. OpenAI's GPT-5.3-Codex is explicitly positioned as moving "beyond code to computer operation." Category boundaries are no longer reliable for procurement decisions.
Finding 2
Who Watches the Agents?
Verification and compliance are emerging as a critical gap. Google's Conductor automates post-implementation code review. Anthropic's Claude Code Security scans for vulnerabilities. Amazon's Kiro caused two AWS outages when operating autonomously. No vendor has a complete answer. Enterprises adopting Level 2-3 agents without corresponding verification infrastructure are accumulating operational risk.
Finding 3
Coding Becomes a Shared Capability
Boris Cherny, head of Claude Code at Anthropic, anticipates coding becoming a shared capability across roles rather than the exclusive domain of software engineers. This is not a 5-year forecast; Cowork already enables non-developers to execute production-quality multi-step workflows. Markets are pricing this as organisational restructuring, not just tool adoption.
Market Landscape

Five Vendor Strategies

Anthropic: Horizontal Agent Platform
One agent architecture (Claude Agent SDK), many surfaces. Claude Code in terminal, Cowork on desktop, Chrome for browser automation, Excel and PowerPoint as Office add-ins. Reported $2.5B ARR for Claude Code alone (Bloomberg). Pure subscriber funding with no vendor sponsorship. The broadest deployment surface.
OpenAI: Model Velocity + Speed
Four Codex model variants in six weeks. Cerebras partnership for 1000+ tokens/second inference. Explicitly rejected MCP for Codex, publishing App Server as alternative architecture. Aardvark security agent and $10M cybersecurity grant programme.
Microsoft: Ecosystem Integration
AI embedded in every M365 application via Copilot. Microsoft Graph for data grounding. GitHub Copilot separately for developers (20M+ users, 90% Fortune 100). $30/user/mo add-on requiring M365 E3/E5. Deepest enterprise integration but most fragmented product surface.
Google: Model + Workspace + Cloud
Gemini embedded in Workspace at aggressive pricing (included from $14/user/mo). Gemini CLI open-sourced. Conductor extension adds automated post-implementation review -- the only vendor systematically integrating verification into the agent workflow. 1M+ token context window.
Open Source / Local-First
OpenCode (95K+ GitHub stars), Ollama (100K+ stars, now Anthropic API-compatible), LM Studio (commercial licence removed). Privacy-first, model-agnostic. 42%+ of developers now running LLMs locally. Kilo CLI ($8M seed, GitLab partnership) signals VC entering the open-source agent space.
Market Dynamics

Consolidation Signals

Consolidation is accelerating. Cognition AI acquired Windsurf (signed Jul 2025, $82M ARR, 350+ enterprise customers) while simultaneously dropping Devin pricing from $500/mo to $20/mo. Cursor acquired code review startup Graphite (500+ enterprise customers including Shopify, Snowflake, Figma). JetBrains shut down its Fleet IDE to pivot entirely to an agentic development product.

Each move narrows the number of independent players and raises the stakes for vendor selection. The window for choosing a "safe" option is closing.

Framework

Four Axes Analysis

Functional
The buying decision is no longer "which model is best" but "which model family covers my range of use cases at sustainable cost." Autonomy level as buying criterion: Level 0-1 is commoditised, Level 2 is where current enterprise value concentrates, Level 3 is where capability and risk are highest.
Application
Enterprise adoption evidence is strong: NVIDIA deployed Cursor to 30,000 engineers (~3x code output), Dropbox reports 90%+ of engineers using AI weekly. But a capability-readiness tradeoff exists: most capable agents have weakest enterprise governance; most enterprise-ready are least capable.
Systems
MCP has achieved near-universal adoption. OpenAI's App Server creates a standards fork. Ollama and LM Studio added Anthropic API compatibility, enabling frontier tools with local models. Integration decisions made in the next 12 months will shape vendor lock-in for 5+ years.
People & Process
Coding as shared capability is not a speculative outlook -- Cowork already enables non-developers to execute production workflows. The $285B stock selloff reflects investor conviction in the displacement thesis. Governance lags capability by 6-12 months (PwC, Gartner data).
Critical Gap

Who Watches the Agents?

Google Conductor (Feb 13 2026) integrates automated review into Gemini CLI workflow. After code generation, Conductor produces reports on quality, style, security (including secrets and PII). The most architecturally integrated approach.

Anthropic Claude Code Security (Feb 20 2026, research preview) scans codebases for vulnerabilities, provides multi-stage verification. Stock market reaction (CrowdStrike -8%, Cloudflare -8.1%) signals investor conviction in displacement thesis.

The Kiro Evidence: Two AWS outages linked to AI agents operating with production permissions: a 13-hour China region outage and a second incident tied to Q Developer. Whether attributed to human misconfiguration or structural agent risk, the incidents demonstrate verification gaps at Level 3 create real-world production failures.

Comparative Data

Pricing Comparison (Feb 2026)

Vendor/Product Individual Team/Enterprise
Anthropic Claude (Pro) $20/mo $25/seat/mo
Anthropic Claude (Max) $100-200/mo Custom
OpenAI Codex Bundled $20/mo; Pro $200/mo Custom
Microsoft 365 Copilot N/A $30/user/mo (req. E3/E5)
GitHub Copilot $10-19/mo $19-39/user/mo
Google Gemini (Workspace) From $14/user/mo Custom
Cursor $20/mo $40/user/mo
Devin From $20/mo Custom
Recommendations

Implementation Guidance by Profile

Microsoft-first Enterprise (M365 E3/E5)
Start with M365 Copilot for breadth, add GitHub Copilot for developer teams, evaluate Claude Code/Cowork for high-autonomy use cases. Monitor Cowork enterprise features for expansion.
Google Workspace-first Enterprise
Gemini Workspace is the natural starting point with aggressive pricing. Add Conductor for verification. Evaluate Claude Code for developer teams where Gemini's coding capabilities fall short.
Developer-heavy Organisation (>30% Engineering)
Evaluate Claude Code and Cursor as primary developer tools. Both have demonstrated enterprise traction. Run a controlled comparison before committing. Watch for Cowork enterprise features.
Regulated Industry (Finance, Defence, Healthcare)
Prioritise M365 Copilot or Google Gemini. Adopt agent coding at Level 2 only. Build verification infrastructure before permitting Level 3 adoption. The Kiro incidents should inform risk acceptance.
Execution

90-Day Evaluation Playbook

Weeks 1-2: Inventory current AI tool adoption (likely fragmented). Map use cases to autonomy levels. Identify high-value Level 2 tasks for pilot.

Weeks 3-6: Run controlled evaluation. Select 2-3 tools (one developer tier, one knowledge-work tier). Measure task completion, quality, time savings against baseline.

Weeks 7-10: Assess integration architecture. Map MCP vs. App Server dependencies. Model 3-year lock-in scenarios. Evaluate verification tooling against risk framework.

Weeks 11-12: Decision brief for leadership. Recommend primary platform, specialist tools, standards posture, governance requirements, workforce planning implications.

Forward Look

What to Watch Next Quarter

Cowork enterprise features (audit logs, compliance API, org-wide plugin management). Availability determines enterprise readiness timeline.

ACP adoption trajectory. If ACP gains traction alongside MCP, multi-vendor interoperability improves substantially.

Kiro follow-up. Whether Amazon publishes a detailed post-mortem will signal industry direction on Level 3 governance.

Conductor adoption. If Google's verification-integrated approach gains traction, expect Anthropic and OpenAI to follow.

Labour-market data. Watch for concrete headcount and role restructuring tied to agent adoption.

Sources

Methodology

This brief synthesises three independently-compiled research artifacts: a coding copilots landscape (435 lines, 100+ sources), an agent taxonomy (403 lines, v2 post-reconciliation), and an external newsfeed (67 citations, independently compiled).

Source quality note: Revenue and valuation figures ($2.5B Claude Code ARR, $1B Cursor ARR, $29.3B Cursor valuation, $285B stock selloff) trace to limited original reporting, primarily Bloomberg and SemiAnalysis. All such figures are reported as attributed claims, not verified facts.