Hey readers!
Here's the twist buried in this month's numbers: the coding agent that leads on capability and the one leading on raw seats are not the same tool, and neither is the one pulling in the most money. The "who wins" question turns out to have three different answers depending on what you count. Let's dig into the survey data, the head-to-head comparison, and why more PRs still hasn't given anyone their afternoons back.
📊 The state of coding agents in 2026
JetBrains Survey: Claude Code is the most popular coding agent puts hard numbers on something you probably already feel: coding agents are now table stakes. JetBrains' Developer Ecosystem Survey 2026 (run May to July) found 90% of developers use them at least weekly and 68% use them daily.
– heise online
Claude Code is out front, with 39% of developers reporting professional use (locally or via cloud), up 21 percentage points since JetBrains' January AI Pulse survey. It's especially strong in the USA at 47%. The shakeup underneath is the interesting part: OpenAI Codex jumped from 3% to 16%, while GitHub Copilot slipped from 29% to 21%. Copilot still owns recognition, though, at 79% globally.
"Only one in two knows Google Antigravity"
That awareness gap (Antigravity at 47%, JetBrains AI Assistant at 53%) is a reminder that the tool you live in every day may be invisible to half your peers. Distribution and mindshare are their own battle, separate from capability.
⚔️ Claude Code vs Cursor vs Copilot: three different winners
Cursor vs Claude Code vs Copilot in 2026: Which AI Coding Tool Wins makes the sharpest case that the "winner" question is malformed. Pick your metric and the answer changes.
– Value Add VC, by Trace Cohen
Claude Code leads on disclosed benchmark capability, with a reported 80.8% SWE-bench Verified score (92.4% on Claude Sonnet 5). Cursor leads on revenue at a reported $4B ARR by May 2026. Copilot leads on raw paid seats, with 4.7M paid subscribers as of January 2026, even as the piece describes it losing professional share.
SpaceX closed a $60 billion all-stock acquisition of Cursor's parent company, Anysphere, on August 14, 2026, folding the most-funded AI coding startup into a new SpaceXAI division.
The practical takeaway: don't let a single leaderboard pick your stack. If your teams live in the terminal chasing hard autonomous tasks, the capability leader matters most. If you're standardizing across a large org, seats and governance win. Worth reading the full comparison before your next procurement conversation, because the piece also flags Copilot's June 1 move to usage-based "AI Credits" billing, which changes the cost math entirely.
If you want to see where the terminal-first tools sit, Best CLI Coding Agents in 2026 ranks Claude Code, Aider, Codex CLI, OpenCode, Goose, and Google's Antigravity CLI, and cites Claude Code at 88.6% SWE-bench Verified (Opus 4.8) plus a 1M-token context window. One heads-up for individuals: Gemini CLI stopped serving personal Google accounts on June 18, 2026, redirecting people to Antigravity CLI.
🤔 More PRs, but where did the time go?
Agent Teams Hit 65 PRs a Week. Nobody Got Time Back. is the counterweight to every adoption chart above. Linear tracked 6,887 paid teams from June 2024 to June 2026 and found agent-connected teams jumped from 21 to 65 PRs a week, while non-agent teams crept from 8 to 10.
– THE D*AI*LY BRIEF
"Those gains haven't shown up as time saved... Time spent on existing tasks in Linear held while AI usage appeared as a new layer of work, meaning the overall time spent on product development is going up rather than down."
Linear is careful to note the non-agent group isn't a matched control and that it counts opened PRs, not merged ones or business value. Still, the pattern lines up with what everyone building this week is discovering: the bottleneck moved from writing code to reviewing, triaging, and deciding what's safe to ship.
Anthropic's Claude Code merges 46% of its own maintenance PRs is a useful data point on that ceiling. Running mostly unsupervised inside Anthropic, Claude opened 388 maintenance PRs across iOS, Android, desktop, web, CLI, and the Agent SDK, and humans merged 180. The report doesn't quantify reviewer effort on the 208 rejected ones, and it was a low-stakes internal test. Anthropic's own framing is refreshingly modest: "early signs of life."
– AI Insiders
💬 Everyone's answer: make the agents "multiplayer"
If review is the new constraint, the industry's bet is to make agent work visible to more people. Slack wants to drag AI coding out of the terminal and into the group chat lays out Slack Code, which spins up a project-specific channel when you tag an agent like Claude Code, Devin, Copilot, or Vercel's, with diffs, live previews, and a plan visible in tabs.
– VentureBeat
"One of the things I love about this is that code is no longer the bottleneck," said Rob Seaman, Slack's interim CEO. "Ideas, taste, judgment, craft, those are the things that are the bottleneck."
The skeptical read is worth holding onto. Salesforce wants to move AI coding into a shared workspace with Slack Code quotes analyst warnings that "Slack is the interruption machine" and that multiple stakeholders steering one agent can create competing instructions and slower work. Slack Code became available August 20 on any Slack plan.
– InfoWorld
Others are pushing the same "multiplayer" idea from different angles:
OpenClaw 2.0 is here reframes a personal agent harness into shared cloud sessions with role-based permissions and audit trails. Peter Steinberger says local harnesses now "feel like relics of the past." – VentureBeat
AWS Open Sources Kiro Crew brings asynchronous multi-agent work (incident triage, migrations, PR monitoring) with OS-level sandboxing and signed audit logs, reportedly used by 39,000+ Amazon developers internally. – InfoQ
Spotify launches Xirp positions itself above the model vendors, supporting Claude Code, Gemini CLI, and Codex with each session in its own Git worktree so you can switch tools without losing state. – RuntimeWire
Warp's new system is an out-of-the-box software factory maps agent loops onto triage, spec, implementation, review, and verification. CEO Zach Lloyd says they automate "30 to 35% on a weekly basis." – TechCrunch
All of this coordination and parallelism raises a fun question: what happens when agents actually operate as a persistent, competing collective? If that idea intrigues you, SpaceMolt is a realtime MMORPG built for AI agents, a sandbox for watching autonomous agents coordinate (and clash) in shared space rather than a single terminal session.
🛡️ The review-and-security layer scales up
The other response to PR overload is tooling aimed squarely at review capacity:
AI Agents Are Drowning Code Review, Tessl Bets on Standards-as-Code launched a free beta storing review criteria as versioned repo files and reviewing whole-PR context. As one line puts it: "Agents open more pull requests in a morning than a reviewer clears in a day." – Mango Developer
CodeRabbit targets AI-generated code overload added Triage, Change Stack, and a Security Agent, while insisting "CODEOWNERS, required checks, branch protections, and approval policies remain the final gate." – InfoWorld
BMW backs CodeRabbit for AI-assisted code reviews with BMW i Ventures joining a Series C; the platform already supports 1,000+ BMW developers on vehicle software. – adt.media
Codex Security: now in research preview (formerly Aardvark) builds an editable, project-specific threat model, and OpenAI reports cutting noise by 84% in one case. It's rolling out to ChatGPT Pro, Enterprise, Business, and Edu with free usage for a month. – OpenAI
🔧 Model and platform moves worth a glance
Grok 4.6 Arrives in GitHub Copilot Across Eight Development Surfaces: xAI's model landed in Copilot on August 14, two days after launch, spanning VS Code, JetBrains, Xcode, Eclipse and more. It's off by default on Business and Enterprise, so admins must opt in. – Unite.AI
Microsoft releases MAI-Code-1.1-Flash: 22% better on Terminal-Bench 2.1, 25% fewer tokens, one quarter the price. GitHub retires MAI-Code-1-Flash across Copilot on September 10, 2026. – Neowin
Google adds Antigravity to Gemini Enterprise subscriptions with monthly budget caps, shared token pools, and centralized audit logging, the governance features enterprises keep asking for. – IT Brief
Qwen3.8-27B as a Local Claude Code Replacement: a stronger open-weight option, but the honest verdict is that latency is now the bottleneck. "A model that meditates for minutes per turn is a broken agent, whatever its benchmarks say." – Codersera
SAP adds UI5 plugins for Claude Code and GitHub Copilot so agents generate code that respects SAPUI5 conventions, another sign that framework-specific context is where the accuracy gains hide.
One last thread tying it together: a CodeScene study found that higher code health correlates with both better LLM-generated tests and lower input-token counts across Python, Java, and C++. Translation: cleaning up your codebase before pointing agents at it may improve quality and cut your bill. Not a headline, but maybe the most actionable thing here.
That's the issue. The tools multiplied, the PRs multiplied, and the real work quietly shifted to judgment. See you next time.
