Why the New Fast‑Agent Era Matters: Cursor 2.0, Composer, Windsurf & SWE‑1.5

Introduction

Kate introduces the newly released Composer model from Cursor and the SWE‑1.5 model from Cognition, highlighting them as a new stage for AI‑assisted programming.

Performance breakthrough

Official benchmark charts show both models surpass the previously celebrated Haiku 4.5 on the SWE‑Bench leaderboard. In the author’s own tests, Composer not only outperforms Haiku 4.5 but does so with considerably higher speed. Cognition reports that SWE‑1.5 processes roughly 950 tokens per second , a very fast rate.

Importance of fast feedback loops

Rapid models enable two to three iteration cycles within a few minutes, which is crucial for tasks such as A/B testing. By contrast, a larger, slower model may take ten minutes for a single prompt. Community member swyx reports that Composer completed a task after two rounds of human feedback, while Sonnet 4.5 was still loading.

Pricing and value

Composer is priced similarly to GPT‑5 Codex, making it not cheap, but its speed and quality constitute its core value proposition.

Development history

Composer was first released in July of the previous year with multi‑file editing capabilities. Since then, many AI coding tools have adopted a similar approach, and the agent capabilities have continued to improve. Composer uses a Mixture‑of‑Experts (MoE) architecture and supports up to 200 K tokens of context . Cursor Bench scores indicate that Composer exceeds the July frontier models, and Cursor explicitly states that Composer’s coding ability surpasses Sonnet 4. Fast frontier models mentioned include Gemini 2.5 Flash, Haiku 4.5, and the open‑source GLM 4.6. The model originated from the Cursor Tab project and a prototype called Cheetah , with Composer being the more intelligent version.

Reinforcement learning power

During training, the Composer team applied reinforcement learning to enable the model to perform complex searches, fix linter errors, and run unit tests. Engineers said RL fine‑tuning was chosen because it most effectively improves the model’s interactive‑agent capabilities. The model is not based on Grok.

SWE‑1.5 technical characteristics

SWE‑1.5 demonstrates very high speed (highlighted in the official chart) and achieves a top score on the new SWE‑Bench Pro benchmark. The model is trained end‑to‑end with reinforcement learning and iterates continuously. It performs best when paired with Windsurf’s Cascade; its score drops significantly when used inside Claude Code. Training took place on new GB200 chip clusters in collaboration with Cerebras, which explains the speed advantage. At such high speeds, system latency became the new bottleneck, prompting a rewrite of the linter and command‑execution components.

Cursor 2.0 major updates

Cursor 2.0 adds a sidebar that displays an agent plan and allows up to eight agents to run in parallel per prompt. Agents operate in isolated Git worktrees or remote machines to avoid file conflicts. The interface is split into Agents and Editor panels. The new browser integration lets users select page elements and converse with them. Code‑review features have been improved, and the default execution environment on macOS runs agents inside a secure sandbox. Users can customize commands and rules, and a voice‑button has been added. Users may plan with one model and build with another, and multiple agents can be run in parallel.

Hands‑on experiment: Periodic‑table project

Using Composer, the author generated roughly 1500 lines of code for an interactive periodic‑table web page. The model ran at about 1.1× the speed of previous attempts, automatically fixed syntax errors, and produced an IUPAC‑compliant layout. Compared with DeepSeek V3.2 Exp (which failed after a few elements), GLM 4.6 (non‑IUPAC), and Haiku 4.5 (slower and error‑prone), Composer achieved the highest speed and completeness.

Browser OS project

Composer built a high‑fidelity web‑OS in stages: it first created a todo list, then fixed calendar bugs, and finally simplified the calendar code. The resulting UI resembled a Windows‑like environment with Wi‑Fi, weather, and time displays. Parallel execution was demonstrated by running Grok Code (free but rate‑limited) alongside Composer, showing both agents operating side‑by‑side.

Digital drawing app project

Composer outperformed Haiku 4.5 by completing a digital‑drawing web app in about one minute. The app includes a pattern library, image‑to‑sketch mode, ink‑like strokes, and playback of the drawing process. The UI responded smoothly to color changes, eraser use, and other interactions.

Fake Monkey King project & parallel agents

In plan mode, Composer quickly generated a plan, then built the project, automatically fixing code issues, updating the todo list, and handling compatibility problems. The parallel agents view displayed Grok Code (free with limits) on the left and Composer on the right, both editing the same repository in separate worktree branches. After completion, the author applied all changes to the main branch.

Cost analysis

The periodic‑table project (≈1500 lines) cost $0.31 per run, while the WebOS project (≈2000 lines) cost $0.29 .

Conclusion

Overall, the Composer model delivers fast, high‑quality coding, and the author looks forward to testing SWE‑1.5. The experience confirms that AI coding tools are moving toward greater speed, intelligence, and practical usefulness, with fast feedback loops becoming increasingly critical.