Agentic Software Engineering
Training and evaluating models that act as agents in real repositories, reading code, running tests and submitting fixes for issues.
Tier Grades
Tier A: Frontier AI labs
feasible
OpenAI, Anthropic, Cognition and Cursor post-train coding agents at scale.
Tier B: Indonesian AI labs
stretch
Fine-tuning a 32B model on 50k trajectories (SWE-smith) is a compute stretch; Indonesian capacity is not in the sources.
Tier C: Institution units such as ITB
stretch
SWE-smith came from a 10-author team, but it builds execution environments for 128 repositories.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Delivery teams can use released coding agents (Devstral, OpenHands LM) on client repositories in proofs of concept.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeSWE-agent-LM-32B is fine-tuned from a 32B open model on SWE-smith trajectories; training compute is not in the abstract.
- DataSWE-smith: 50k task instances from 128 GitHub repositories, an order of magnitude larger than earlier work; SWE-agent-LM-32B reaches 40.2% Pass@1.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 10 authors on SWE-smith.
- Time horizonMonths.
- Shelf-life signalsTool vendors post-training their own coding agents on open bases (Cursor Composer, Cognition SWE records in AURION), SWE-bench style leaderboards changing within months
- Pipeline roleTier A trains agents and benchmarks; Tier C builds Indonesian repository tasks; Tier D uses agents to speed client software delivery.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Alibaba (Qwen)
OpenAI
Mistral AI
Cognition (owns Windsurf)
Anysphere (Cursor)
All Hands AI
Anthropic
Google DeepMindLinked models
Devstral Small 2 (24B, 2512)codeexemplarGPT-5.1-Codexcodeexemplar
SWE-agent-LM-32BcodeexemplarOpenHands LM 32B v0.1codeexemplar
MiMo-V2.6-Flashmultimodal LLMproduced bySWE-2codeproduced byGemini 3.8 FlashLLMproduced byQwen3.8-27Bmultimodal LLMproduced byGemini 3.6 FlashLLMproduced by
Laguna S 2.1codeproduced by
Muse Spark 1.1LLMproduced by
North Mini Codecodeproduced byComposer 2.5codeproduced byQwen3.6-35B-A3BLLMproduced by
GLM-5V-Turbomultimodal LLMproduced byQwen3.6 Max PreviewLLMproduced byQwen 3.6 PlusLLMproduced byComposer 2codeproduced byGPT-5.3-Codexcodeproduced byQwen3-Coder-Nextcodeproduced byGPT-5.1-Codex-Maxcodeproduced byDevstral 2 (123B, 2512)codeproduced byGPT-5.1-Codex-Minicodeproduced by
KAT-Coder-Pro V1codeproduced by+17
grok-code-fast-1codeproduced byQwen3-Codercodeproduced byDevstral Medium 1.0codeproduced byDevstral Small 1.1codeproduced byQwen3-Coder-30B-A3B-Instructcodeproduced byQwen3-Coder-480B-A35B-Instructcodeproduced bySWE-Gym agentscodeproduced byClaude 3.5 Sonnet (Oct '24)multimodal LLMproduced byClaude Mythos 5LLMproduced byComposercodeproduced bySWE (Devin)codeproduced bySWE-1.6 previewcodeproduced bySWE-1.7codeproduced by
MAI-Code-1.1-Flashcodeproduced byDevstralcodeproduced bycodex-mini-latestcodeproduced byGPT-5.2-Codexcodeproduced byLinks
R&D TopicAgentic Software Engineering
ModelsOpenHands LM 32B v0.1All Hands AI, codeDevstral Small 2 (24B, 2512)Mistral AI, codeGPT-5.1-CodexOpenAI, codeSWE-agent-LM-32BSWE-bench team (Princeton/Stanford), code+37 more
OpenHands LM 32B v0.1All Hands AI, codeDevstral Small 2 (24B, 2512)Mistral AI, codeGPT-5.1-CodexOpenAI, codeSWE-agent-LM-32BSWE-bench team (Princeton/Stanford), code+37 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 41
Claude Mythos 5Anthropic, LLMproduced byDevstralMistral AI, codeproduced byGemini 3.6 FlashGoogle DeepMind, LLMproduced byMiMo-V2.6-FlashXiaomi, multimodal LLMproduced byClaude 3.5 Sonnet (Oct '24)Anthropic, multimodal LLMproduced byComposer 2.5Anysphere (Cursor), codeproduced byGemini 3.8 FlashGoogle DeepMind, LLMproduced byGPT-5.1-CodexOpenAI, codeexemplarMuse Spark 1.1Meta, LLMproduced byQwen3-CoderAlibaba (Qwen), codeproduced bySWE (Devin)Cognition (owns Windsurf), codeproduced bySWE-1.7Cognition (owns Windsurf), codeproduced by+29
SWE-2Cognition (owns Windsurf), codeproduced byComposerAnysphere (Cursor), codeproduced byComposer 2Anysphere (Cursor), codeproduced byDevstral Small 2 (24B, 2512)Mistral AI, codeexemplarGLM-5V-TurboZ.ai (Zhipu), multimodal LLMproduced byGPT-5.1-Codex-MaxOpenAI, codeproduced byGPT-5.2-CodexOpenAI, codeproduced byQwen 3.6 PlusAlibaba (Qwen), LLMproduced byQwen3.6 Max PreviewAlibaba (Qwen), LLMproduced byQwen3.6-35B-A3BAlibaba (Qwen), LLMproduced byQwen3.8-27BAlibaba (Qwen), multimodal LLMproduced bySWE-1.6 previewCognition (owns Windsurf), codeproduced bySWE-agent-LM-32BSWE-bench team (Princeton/Stanford), codeexemplarDevstral 2 (123B, 2512)Mistral AI, codeproduced byDevstral Medium 1.0Mistral AI, codeproduced byDevstral Small 1.1Mistral AI, codeproduced byGPT-5.1-Codex-MiniOpenAI, codeproduced byGPT-5.3-CodexOpenAI, codeproduced bygrok-code-fast-1SpaceXAI (xAI), codeproduced byKAT-Coder-Pro V1KwaiKAT, codeproduced byNorth Mini CodeCohere, codeproduced byOpenHands LM 32B v0.1All Hands AI, codeexemplarQwen3-Coder-30B-A3B-InstructAlibaba (Qwen), codeproduced byQwen3-Coder-480B-A35B-InstructAlibaba (Qwen), codeproduced byQwen3-Coder-NextAlibaba (Qwen), codeproduced bycodex-mini-latestOpenAI, codeproduced byLaguna S 2.1poolside, codeproduced byMAI-Code-1.1-FlashMicrosoft AI, codeproduced bySWE-Gym agentsAll Hands AI, codeproduced byPapers
- SWE-bench: Can Language Models Resolve Real-World GitHub Issues?Carlos E. Jimenez et al., ICLR 2024 (per arXiv comment), 2023 founding
- SWE-smith: Scaling Data for Software Engineering AgentsJohn Yang et al., arXiv preprint, 2025 state of the art
- AutoCodeRover: Autonomous Program ImprovementYuntong Zhang et al., arXiv preprint, 2024 state of the art
- SWE-agent: Agent-Computer Interfaces Enable Automated Software EngineeringJohn Yang et al., arXiv preprint, 2024 state of the art
- Large Language Model-Based Agents for Software Engineering: A SurveyJunwei Liu et al., arXiv preprint, 2024 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2504.21798 https://arxiv.org/abs/2504.21798
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Agentic Software Engineering", https://aurion.semai.id/topics/agentic-software-engineering/, records of October 2026.