New Qoder: AI Coding's Exit Is Not Inside the IDE
One-line summary: On August 27, Alibaba rebuilt Qoder from an "AI coding IDE" into an "agent workbench with coding as core capability, for everyone." A year of accumulated Agent Harness work now pairs with 40+ connectors, 70+ plugins, and 20,000+ skills, plus a three-layer long-horizon structure (Plan / Goal / side tasks) and a desktop pet with real-time voice entry — covering 6 million users and 100,000 enterprise customers. In the same week, Claude Code shipped --restricted, GitHub Copilot filled in Customize and enterprise model policies, and Huawei's CodeArts Agent went commercial in Asia-Pacific. The competitive axis of AI coding is migrating from "can the model write correct code" to "how many real work systems are connected, and can the agent be governed."
Key points
1. Three shifts away from the editor
- Work object: the entry point changes from a file/repo to a natural-language goal ("analyze this module's refactoring plan first"). Code recedes to the background execution layer.
- Work mode: from personal production to delegation and acceptance — you give the agent a goal, it runs, you interject and verify.
- Participants: a "general mode" explicitly targets product, ops, data, design, and indie creators — people who can judge output quality but don't write code.
- Cross-session long-term memory — no re-explaining context from scratch
- Automatic context compression — remembering step 3 at step 50
- Periodic skill refinement — high-frequency success paths become reusable skills
- Security: tiered permissions and tool whitelists; sensitive operations require explicit authorization. This is the precondition for delegation to be viable.
- Aug 25 — GitHub Copilot's Customize tab GA: MCP servers, plugins, skills, canvas in one place (ecosystem layer).
- Aug 26 — Copilot global model policy GA: admins set defaults; open-weight and data-retention model classes off by default (governance layer).
- Aug 27 — Copilot code review adds "addressed reason" labels (Addressed / Won't fix / Incorrect) and extends coverage to bot PRs, cloud-agent PRs, and large PRs beyond the prior 300-file / 20,000-line ceiling (audit layer).
- Aug 27 — Claude Code v2.1.248 ships
--restricted, an out-of-the-box sandbox for CI and untrusted contexts (sandbox layer); v0.150.1 (Aug 26) fixed image tokens counted against budget during remote compression. - Aug 28 — Huawei Cloud CodeArts Agent GA in Asia-Pacific: 16 built-in professional agents covering requirements, architecture, coding, testing, debugging, and review; 30+ reusable engineering skills; claims understanding of codebases at the ten-million-line scale (engineering-skills layer).
> Note: An Agent Harness is the engineering skeleton wrapped around the model — context assembly, tool calling, retry logic, long-session compression. The model is the engine; the harness is the chassis and transmission. The same model with a different harness can differ by more than 2x in results.
2. Read — Edit — Verify — Iterate
The new Qoder's technical base is a one-year-refined Harness summarized in four verbs: read (retrieve the codebase) → edit (cross-file) → verify (run tests) → iterate (self-correct from real results). Many tools stop at read + edit, leaving humans to catch compile/test failures. Compilers and tests are judges that don't lie — they return 0/1 signals.Supporting engineering pieces:
3. Ecosystem: 40 + 70 + 20,000
| Layer | Count | Role | Analogy | |---|---|---|---| | Connector | 40+ | Plug in code repos / project management / cloud services / internal tools | Utilities hookup | | Plugin | 70+ | Extend concrete executable actions | Appliances on the socket | | Skill | 20,000+ | Encapsulate stable standard workflows | A written SOP manual |
With Browser Use / Computer Use support, agents can operate web pages and desktops directly — "old systems without APIs" are no longer blind spots. For enterprises, the most expensive part of a task isn't writing code; it's stitching context scattered across a dozen systems.
4. Auto routing and model–harness co-optimization
Built-in frontier models including Qwen3.8-Max plus Auto intelligent scheduling balance quality, speed, and cost per task. Officially, the Harness and Qwen models iterate jointly — end-to-end optimization on long context, compression, and tool-call accuracy. This mirrors Anthropic's Claude Code and OpenAI's Codex: first-party model + first-party harness vertical integration is displacing "generic model wrapped in a third-party shell."5. Desktop pet + real-time voice: an underrated interaction signal
A resident desktop pet accepts spoken requests with live progress sync, alongside a scratchpad, meeting notes, and real-time voice discussion. It solves a real delegation problem: watching the agent negates delegation; ignoring it leaves no one to catch errors. The pet offers "low-bandwidth residency + interrupt anytime." Official example: a user says "analyze this module's refactoring plan," Qoder generates a plan from project context; the user adds "add caching" by voice, and execution/verification continues — no need to restate complete instructions each time. Text prompts suit batch processing; voice suits long-horizon interactive iteration.6. Product matrix: seven form factors
| Form | Scenario | |---|---| | New Qoder (desktop) | Main workbench, natural-language entry | | Qoder IDE | Professional developers | | Qoder CLI | Scripted / automated pipelines | | Qoder JetBrains plugin | Existing IDE users | | Qoder mobile | Light viewing and tasks | | QoderWake | Digital-employee form | | Qoder Cloud Agents | Continuous cloud execution |
An agent's value density depends on how many real workflows it can appear in — from requirement to delivery, not just "writing code."
7. A five-dimension evaluation framework
| Dimension | Question to ask | Failure mode | |---|---|---| | Context integration | Can it really read repos, requirement systems, cloud services, internal tools? | Only reads the current file; humans paste the rest | | Long-horizon stability | Does it remember step 3's constraints at step 30? | Drifts further until delivery misses the target | | Verification loop | Self-corrects, or dumps raw compile errors on you? | Only read + edit, no verify + iterate | | Governance & permissions | Can you limit which files and tools it touches? | All-open permissions; enterprises won't adopt | | Cost predictability | What does a medium task cost? | One refactor burns a month's budget |Long-horizon stability shows the widest gap between Codex, Claude Code, and the new Qoder — it decides whether an agent is "usable for ten steps" or "usable for an afternoon."
8. The same week: the whole lane shifts gears
9. A judgment worth cautioning
The new Qoder narrative's most striking line: letting non-technical staff build prototypes and validate ideas without waiting on engineering schedules. If agents compress weeks of scheduling to hours, organizational innovation rates change by orders of magnitude. But the distance between prototype and product is determined by reliability, maintainability, and security — not code volume. A collapse under real traffic is often paid for by the very engineering team that was bypassed. The likelier evolution: everyone runs a first version cheaply, and professionals decide what deserves engineering. The most valuable node in that loop is cheap rejection — how fast an organization kills bad ideas often matters more than how fast it implements good ones.10. Three observable windows over the next six months
1. General-mode retention. Whether non-technical users stay is the only hard test of whether "coding capability spilling over" is real — otherwise this is just a UI refresh. 2. The long tail of 20,000 skills. Skill counts are easy to inflate; if 90% of calls concentrate in 100 skills, the number is inventory, not capability. 3. Auto's actual cost savings. If Auto cuts task cost to under a third of "always use the strongest model," model routing becomes a core moat; if not, it's a marketing word.Closing analogy: After Watt's steam engine patent in 1769, power was no longer scarce — what was scarce was the transmission that connected power to looms, mine carts, and ships. AI coding is in the same position: model "power" is abundant; the transmission connecting it to real work systems is scarce. 40 connectors, 70 plugins, 20,000 skills — Alibaba has submitted its version of the transmission. Whether a transmission is good isn't judged by specs, but by how many machines it can drive.