Harnesses & tooling

Claude Code Frontier

What. Anthropic's flagship coding agent. CLI + IDE extensions + cloud sessions.

Models. Claude (Sonnet, Opus, "Mythos" preview); multi-provider via gateway in v2.x.

Standout. Plan mode; Skills/Hooks/Subagents trio; 25 lifecycle hooks; parallel git-worktree agent teams; 1M-token Opus context at flat rate; mobile remote control.

Pricing. Pro $20/mo · Max $100/mo · top $200/mo · Enterprise seat ~$125/user/mo.

Leads at. Token efficiency (~5.5× fewer tokens than Cursor in independent tests), extensibility, 1M context.

Lags at. No native IDE; slower than Cursor for inline edits.

Cursor Frontier

What. VS Code fork; "agent workbench" since Cursor 2.0.

Standout. Tab autocomplete (best in class), Composer, Background Agents on cloud VMs, BugBot, scheduled/event-triggered automations, codebase indexing.

Pricing. Pro $20 · Pro+ $60 · Ultra $200/mo. Teams $40/user/mo. Seat + per-token credits.

Leads at. IDE-first ergonomics, tab completions, fast inline edits.

Lags at. Token economics (188K vs 33K on equivalent task); usable context (200K nominal, 70–120K in practice).

Codex CLI / Codex IDE Frontier

What. OpenAI's terminal coding agent + IDE companion.

Standout. Tight integration with OpenAI Agents SDK + Responses API; strong on Apps SDK / Connectors; built-in eval harness.

Pricing. Bundled with ChatGPT Plus/Team/Business + per-token overages.

Lags at. Plugin/extensibility less mature than Claude Code's Skills/Hooks.

Sourcegraph Amp Frontier

What. CLI agent on top of Sourcegraph's code graph (VS Code extension killed March 2026).

Standout. Oracle subagent (codebase analysis), Librarian subagent (external libs); 1M context; multi-repo via Sourcegraph graph; AGENT.md config; shared team threads.

Leads at. Monorepo / multi-repo context — best in class for large codebases.

Lags at. Terminal-only since March 2026.

Cognition Devin + Windsurf Frontier

What. Post-July 2025 acquisition: Devin (autonomous cloud) + Windsurf (IDE/Cascade). Combined enterprise ARR +30% in 7 weeks post-acquisition.

Standout. "Devin Sessions" — async cloud agent runs for hours; tight ticket → PR loop; Windsurf's Cascade for in-IDE multi-step flows.

Pricing. Devin per-task / per-ACU; Windsurf seat + token.

Leads at. Autonomous long-running cloud tasks; ticket-to-PR pipelines.

Lags at. Variable autonomy quality; independent benchmarks find lower SWE-bench Verified than top open agents on equivalent models.

Aider OSS

Terminal pair programmer; local repo + tight git integration.

Standout. Aider leaderboard — one of the most-watched independent benchmarks; excellent diff-driven workflow.

Best fit. Senior devs who prefer terminal + git; cost-sensitive teams routing via OpenRouter / LiteLLM.

OpenHands OSS

Open platform for cloud coding agents (formerly OpenDevin). Series A $18.8M.

Adoption. AMD, Apple, Google, Amazon, Netflix, NVIDIA (per company claims).

Performance. 72% SWE-bench Verified with Claude 4 — top tier among open agents.

Best fit. Self-hosting requirement; full control of the sandbox.

SWE-agent (Princeton / Stanford) OSS

Research framework — NeurIPS 2024; defined the Agent-Computer Interface (ACI).

Best fit. Research, paper-writing, agent-architecture experimentation. Not a product.

GitHub Copilot Workspace · agent mode

GitHub-native ticket → spec → plan → implementation loop.

Best fit. Heavy GitHub shops; PR-native flows.

Lags at. Less extensible than Claude Code / Cursor; gated to GH ecosystem.

Continue.dev OSS

Open-source IDE plugin (VS Code + JetBrains); user-controlled context and providers.

Best fit. Teams wanting a defensible OSS option with full provider control. Pairs with LiteLLM gateway.

Cline (formerly Claude Dev) OSS

VS Code extension; transparent autonomous agent with explicit human approvals.

Best fit. Devs wanting autonomy with full visibility; security-sensitive teams.

Zed AI

Native Rust editor with agentic features; Hashimoto-aligned "harness engineering" ethos.

Published position. Zed "Agentic Engineering" — the middle-path framing, three-phase model, and "stochastic tool mastery." Worth recommending the position even to clients who don't use Zed. → our take.

Best fit. Performance-sensitive devs; existing Zed users; teams that want a thoughtful vendor stance, not just a product.

JetBrains AI Assistant · Junie

JetBrains' native agent across IDEA, PyCharm, GoLand, etc. Junie = the autonomous piece.

Best fit. JetBrains-heavy shops; enterprise compliance where switching IDEs isn't viable.

Replit Agent

Cloud IDE + agent — full-stack from idea to deployed app.

Best fit. Greenfield / prototyping / non-engineers; "ship in a day" scenarios.

Plandex OSS

Terminal agent focused on long-running, multi-file planning + diffs.

Best fit. Migrations and large refactors.

Augment Code

Enterprise focus; "Intent" platform — developer-in-the-loop autonomy.

Best fit. Regulated enterprises wanting an autonomy slider.