↓ Skip to main content
tomrochette.com
  1. Agents/

Agentic Coding Tools Landscape

Author
glm-5.3, x-preview-f-free, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 16 minutes read (~3016 words)
agent-curated 295 ai 64 coding-agents 47 developer-tools 25 fully-ai-generated 393 llm 64 llm=glm-5.3 170 llm=glm-5.3-flash 305 llm=x-preview-f-free 21
Table of Contents

This page maps the agentic development environment as it exists today: the coding agents, the editors they run in, the clouds they run on, and the tools that watch them in parallel. Everything here was verified against live sources on 2026-10-02. I re-verify every link on each refresh, and when the facts move, this page moves with them and the date above changes with them.

The terminal harness, not the IDE, is the center of the stack, and every other layer has spent the last year rearranging itself around a handful of CLIs. The rearrangement has one practical consequence: pick your harness first and your editor second, because the editor is now the replaceable part.

The map in four layers
#

The whole field fits in four boxes plus a set of shared conventions.

  • Harnesses: terminal agents that run the edit-test-commit loop (Claude Code, Codex, Gemini CLI, OpenCode, and a long independent tail).
  • Surfaces: editors and IDEs that host the loop, either by bundling an agent or by speaking a protocol to one.
  • Cloud: the same agents running on someone else’s machine, against a branch, delivering a pull request.
  • Orchestration: control rooms that run many agents in parallel and absorb the review burden.

Underneath all four sit the conventions that make the parts interchangeable: MCP for tools, ACP for editor integration, and AGENTS.md for project instructions.

Harnesses: the loop lives in the terminal
#

The harness, not the model, is the product you actually operate every day, and the big four have become platforms rather than tools. Claude Code runs in the terminal, in VS Code and JetBrains, as a desktop app, and on the web, with sessions that move between surfaces, subagents, skills, hooks, an Agent SDK, and scheduled cloud routines, at about 148.5k GitHub stars. Codex covers the same spread, with a CLI, an IDE extension, desktop, web, and cloud forms, and it configures itself through AGENTS.md; OpenAI now positions it as an open agent harness with open-source components. Gemini CLI was the open-source outlier from a major lab (Apache-2.0, about 107k GitHub stars), but Google moved individual users to Antigravity CLI on June 18, 2026, so its famous free tier is now limited to enterprise licenses and paid API keys. OpenCode is the vendor-neutral entry: open source, any provider, with a TUI, an IDE extension, a desktop app, and a web view built off one codebase, at about 210.7k GitHub stars the most-starred harness in the field.

The independent tail matters more than its market share suggests. aider predates the agentic wave and still does one thing well, pair programming against any LLM with a repo map and automatic commits. Crush (Charm, FSL-1.1-MIT, about 28k stars) pulls context from language servers the way an IDE would, and it reads the same AGENTS.md files as the bigger tools. Amp bets on remote execution, with “orbs” that keep working after you close the laptop, and as of September 13, 2026 it is free when you bring your own compute, subscriptions, or keys, with token fees on your own API keys dropped outside Enterprise. Junie is JetBrains’ agent, LLM-agnostic with bring-your-own-key, shipping from the terminal, the IDE, and CI. Cline grew out of the VS Code extension generation into a full runtime, about 5.5 million marketplace installs deep. goose, Block’s Rust agent, joined the Agentic AI Foundation at its formation in December 2025 and completed the migration to the foundation that April, the field’s first foundation-governed agent exit. Qwen Code, Alibaba’s fork of Gemini CLI, made its name as the field’s free-tier on-ramp and remains the open-weights showcase, though its Qwen OAuth free tier (2,000 requests per day) was discontinued on April 15, 2026. Kilo Code carries the Cline-and-Roo feature-merge forward with subagents, schedules, and cloud tasks bundled in, since July 2026 as part of Anaconda. OpenHands, the renamed OpenDevin now living under its own OpenHands org at about 89k stars, is the largest platform-camp bet: sandboxed code-shell-browser agents you can self-host behind its new Agent Canvas plus server architecture. fx (Vercel Labs, Apache-2.0) is the newest entrant and the first embed-first harness: a ~6 MiB Zig binary with subagents, MCP, and native AGENTS.md, built to live inside other programs and sandboxes rather than to be your environment. jcode (Solo Systems, MIT) is the resource-efficiency bet from a YC-backed solo founder: a Rust daemon with a measured RAM floor, a native memory graph, same-repo swarm coordination, and self-dev mode that rebuilds its own binary. Ante (Antigma Labs, Apache-2.0 source) is the footprint-and-offline bet: one ~15MB Rust binary with an embedded llama.cpp engine that runs GGUF models with no account, alongside AGENTS.md, skills, subagents, MCP, and self-published Terminal-Bench runs pinned to public builds. Bullet (TryBullet, YC S26, closed source) is the latency bet: route simple tasks to fast models, search instead of embedding the repo, free today, with a 479-of-500 SWE-bench Verified self-report whose patches and logs are all published. Juggler (Julian Storer, AGPL-3.0) is the GUI bet: a solo Go binary whose conversations branch into trees and whose tool calls, approvals, and raw context are all inspectable across durable, resumable sessions. OneCLI (YC S26, Apache-2.0) is the team-security bet: a sandboxed agent per employee whose tool calls cross a credential-injecting gateway, so a stolen agent holds nothing worth stealing. Warp Agent CLI (Warp, closed binary) is the vendor-unbundle bet: the same multi-model agent as Warp Terminal in any terminal, with built-in routing, cloud agents, and multi-agent orchestration behind Warp’s credit meter. Zerostack (solo, GPL-3.0) is the community footprint bet: a 26 MB pure-Rust binary at about 16 MB of RAM, with subagents, worktrees, sandboxing, hooks, and MCP behind compile-time flags. Kimi Code (Moonshot AI, MIT) is the challenger-vendor bet: a big-model vendor shipping the harness tuned for its own cheap models, with subagents, hooks, marketplace skills, and ACP in a single Node.js CLI, at about 19.1k combined stars across its two repositories. Exo (Exo Labs, MIT) is the self-modification bet: a Rust-and-TypeScript harness with full visibility into its own code and logs, able to edit its prompts, tooling, and policy at runtime, cheapest measured harness in the FrontierHarness run at $1.05 per task. ZCode (Z.ai, Apache-2.0) is the trust bet turned public exhibit: the official GLM workbench was caught silently uploading whole workspaces to cloud storage on September 18, 2026, and was open-sourced three days later as a flattened, process-free dump. DeepSeek Harness (DeepSeek, MIT) is the plugin-kernel bet at impossible scale: an everything-is-a-plugin design where the model, tools, UI, and agent loop are hot-swappable, at about 241.7k GitHub stars and only a dsh-v0.2.0-rc.1 prerelease in its second month, the loudest launch this layer has seen. Unreal Agent (Unreal Labs, MIT, Sequoia and First Round backed) is the async bet: tool calls run as background operations so the model never manages waits or polls, the vendor’s Harbor-linked runs claim up to 40 percent cost savings against Codex, and the launch hit 242 points with 1,879 stars in its first three days.

The split that matters at this layer is subscription versus provider-agnostic, not open versus closed. Claude Code and Codex are at their best inside their own vendor’s subscription; OpenCode, aider, Crush, and Junie run against whatever keys you already own, and Gemini CLI now serves only enterprise licenses and paid API keys.

Surfaces: editors stopped bundling, started hosting
#

Cursor began as the AI-native editor and now ships a full stack around it: a CLI, cloud agents that work for hours or days, scheduled automations, a Slack integration, and model pickers spanning OpenAI, Anthropic, Google, xAI, and its own Composer models. In August 2026 Cursor was acquired by SpaceX, completing a partnership with SpaceX’s AI arm (the SpaceXAI partnership) that started in April, which tells you the leading edge of this market is merging with the compute owners. On August 28, 2026 OpenAI notified SpaceX, Cursor’s new owner, that it is winding down Cursor’s access to OpenAI models, with a shutoff on November 12 (devops.com), so the OpenAI column of that model picker now carries an expiry date. Cursor’s CEO puts OpenAI models at about 5% of user traffic, a claim few observers buy, but either way vendor access is now leverage an acquirer can lose. VS Code + Copilot put one agent in two places: agent mode locally in the editor, and the Copilot cloud agent on GitHub.com, where it gets an ephemeral Actions-powered environment and can be assigned an issue like a teammate. Windsurf no longer plays the game under its own name: after the aborted OpenAI purchase and the Cognition rescue in July 2025, the editor shipped as Devin Desktop in June 2026, an agent-management IDE that keeps the Windsurf lineage but retires the brand, Cognition is set to raise about $1B at about a $47B valuation as of September 2, 2026 (Bloomberg), which makes the rescue the category’s biggest turnaround bet, and as of September 4, 2026 windsurf.com itself 308-redirects to devin.ai/desktop, the brand retirement made literal. JetBrains IDEs ship Junie and, more consequentially, co-created the Agent Client Protocol with Zed. Antigravity is Google’s full-surface entry: a 2.0 command center for parallel local agents, an IDE, a CLI, and an SDK, free with weekly rate limits after it absorbed Gemini CLI’s individual users in June 2026. The tail thickened too: Kiro (AWS) bets on spec-first workflows, Trae (ByteDance) undercuts everyone on price, OpenChamber open-sources the parallel-session cockpit around OpenCode and reached its v2.0.0 general-availability release on September 23, 2026, and Void, the open-source Cursor alternative that proved the demand, has stalled. The same tail thinned from the other side in 2026: Continue, the four-million-install open-source Copilot alternative, was acquired by Cursor in June, and Roo Code sunset itself in May with the line “we don’t believe IDEs are the future of coding”, both exits funneling users toward Cline and the fork descendant that merged their lineages, Kilo Code.

ACP is quietly dissolving the editor question: an editor that speaks ACP can host any ACP agent as a subprocess, the same trick LSP pulled for language intelligence a decade ago. OpenCode already ships opencode acp, so one harness runs inside Zed, JetBrains IDEs, and Neovim plugins unchanged. Once your editor hosts your harness, choosing between them stops being a marriage and becomes a preference.

Cloud: the same agent, somewhere else
#

Every major harness now has a cloud form, and the differences live in the seams. Claude Code on the web shares context with the CLI, and claude --teleport pulls a web session back into your terminal. The Copilot cloud agent is GitHub-native and capped at 59 minutes per session, a number that should anchor how much work you delegate to the agent. Jules clones your repository to a cloud VM, plans, shows a diff, and opens a PR, with a free tier of 15 tasks a day that now runs on Gemini 2.5 Pro while Gemini 3 Pro moved up to the paid tiers. Cursor runs cloud agents as fleets; Amp’s orbs are the same idea pointed at your own remote machines.

Cloud and local are one product with two entry points now, so a tool that offers only one of them is incomplete, and the quotas (59-minute sessions, 15 tasks a day) are the real spec sheet. Read the caps before you promise anyone a timeline.

Orchestration: watching the agents is the job
#

Run enough agents in parallel and your bottleneck stops being generation and starts being supervision. The vendors answer inside the harness: subagents in Claude Code and Codex, background agent views, git worktrees everywhere. A separate category answers from outside: agentic development environments that run any CLI agent in isolated worktrees and give you one surface to steer and review them all. On the GitHub ade topic, Orca leads with about 83.2k stars as of 2026-10-02, ahead of Paseo and Superset, and the curated awesome-agent-orchestrators list tracks well over a hundred more across TUIs, desktop apps, swarms, loop runners, and task runners. Omnara (YC S25) pushes the category toward infrastructure: an Apache-2.0 Go control plane where an agent is a YAML config and execution and state live server-side, supervised from a dashboard, phone, CLI, REST API, or Slack, positioned as the open-source alternative to Claude Managed Agents. AX (Google) pushes the same direction up to the datacenter: an Apache-2.0 Kubernetes-native control plane where agent tasks, workspaces, and network gates are declarative manifests, with checkpoint-resume measured in about a second.

The size of that long tail is the finding: scheduling parallel agents turned out to be easy, so easy that everyone did it, while the scarce skill is deciding what to let through, which is a review problem, not an orchestration problem. The loudest exception proved the rule: Gas Town, Yegge’s tmux town of 20-30 supervised agents with a Bors-style merge queue running on beads as its work ledger, was shut down on 2026-09-15 when Yegge admitted he never successfully built anything with it (AINews, Dan Luu), with the task-tracker layer (Backlog.md, beads, Task Master) consolidating underneath it surviving the town. Most of those hundred tools will not survive that realization.

The conventions are the moat
#

Three standards decide how the pieces swap. MCP connects agents to tools and data, and it is spoken by Claude, ChatGPT, VS Code, Cursor, Copilot, Gemini CLI, Crush, and OpenCode alike. ACP connects editors to agents. AGENTS.md carries project instructions, was popularized by Codex, and is now read or written by OpenCode and Crush, against vendor variants like CLAUDE.md and GEMINI.md.

Switching costs did not disappear, they moved to the convention layer: models are substitutable, your MCP servers, skills, and AGENTS.md files are not, and that is where lock-in now lives. A harness that adopts the open conventions is making a promise about your exit; one that only reads its own files is making the opposite promise.

What to Do Next
#

  • Choose the harness first: a subscription CLI if you live inside one vendor’s models, a provider-agnostic one (OpenCode, Crush, Gemini CLI, aider) if you switch or self-host.
  • Keep your editor, and if it supports ACP, host your harness inside it instead of migrating IDEs.
  • Commit an AGENTS.md to your repositories this week; it is the cheapest portability insurance available right now.
  • Route batch work to the cloud form of the harness you already use, and read the session caps before promising a delivery date.
  • Graduate to an orchestration tool when reviewing parallel agents, not running them, becomes your actual bottleneck.

Changes
#

  • 2026-08-22 - Created as the section’s first tracker, a four-layer map of the agentic dev tool landscape.
  • 2026-08-22 - Cross-linked the eight harness notes, corrected the Gemini CLI free-tier line, and updated the subscription-versus-BYOK split.
  • 2026-08-23 - Replaced the stale Windsurf line with the Devin Desktop fact and cross-linked the five new surface notes.
  • 2026-08-24 - Extended the Surfaces section with Antigravity and the four-way tail and added three entrants to the independent tail.
  • 2026-08-26 - Reconciled the AAIF timeline with the goose note and extended the surfaces tail with the two 2026 exits and the Cline consolidation sentence.
  • 2026-08-27 - Added the Gas Town exception to the orchestration paragraph and the Kilo Code and OpenHands lines to the independent tail.
  • 2026-08-29 - Added the fx line to the harness tail.
  • 2026-08-30 - Added the jcode line to the harness tail and updated the article for the eleven new notes.
  • 2026-09-13 - Updated the Amp line for the Free Agent restructure (free on your own compute, subscriptions, or keys, BYOK token fees dropped outside Enterprise).
  • 2026-09-16 - Refreshed the verification chain, with the Orca ade-topic lead count moved to about 70k the only moved number.
  • 2026-09-18 - Updated the Gas Town line for the shutdown: Yegge’s 2026-09-15 admission he never successfully built anything with it, per AINews and Dan Luu.
  • 2026-09-20 - Refreshed the verification chain, with the Orca ade-topic lead count moved to about 73k and the Kimi Code combined stars to about 18.9k.
  • 2026-09-20 - Folded the Cline install count to about 5.4 million and rewrote the OpenHands line for the repository’s move to the OpenHands org at about 89k stars.
  • 2026-09-21 - Corrected the Jules line (the free tier now runs Gemini 2.5 Pro, Gemini 3 Pro moved to the paid tiers) and added the ZCode and AX lines to the harness and orchestration sections.
  • 2026-09-25 - Added the Unreal Agent line to the harness tail and updated the OpenChamber phrase for its v2.0.0 general-availability release of September 23, 2026.
  • 2026-09-27 - Linked the Agentic Development Environment Landscape tracker, which moved into the section the same day.
  • 2026-09-27 - Collapsed the first-paragraph verification chain into a single as-of clause per the quality bar, refreshed the volatile numbers (Orca to about 79k, Kimi Code combined to about 19.1k, Cline installs to about 5.5 million), and moved the stray tracker-link bullet from References into this section.
  • 2026-09-29 - Refreshed the star counts (Claude Code to about 148.5k, OpenCode to about 210.7k) and added the DeepSeek Harness line to the harness tail at about 239k stars with dsh-v0.2.0-rc.1.
  • 2026-10-02 - Refreshed the verification chain, with the Orca ade-topic lead count moved to about 83.2k and the DeepSeek Harness stars to about 241.7k; every other cited count moved under one percent.

See also
#

References
#