Best coding agent for Go and Python on DeepSWE (2026)
CodeAF solved the most Go (21 of 34) and Python (20 of 34) DeepSWE tasks of ten harnesses on one model. Results by language, with the limits.
On DeepSWE's Go tasks, CodeAF in senior-dev mode solved 21 of 34, the most of ten harnesses in AgentField's comparison, ahead of mini-swe-agent and Codex CLI at 17 each. On the Python tasks it solved 20 of 34, ahead of mini-swe-agent (18) and Codex CLI (17). Every harness ran the same model, deepseek/deepseek-v4-flash-0731, for one pass of the full set.
At a glance
| Harness | Go (34) | Python (34) | TypeScript (35) | JavaScript (5) | Rust (5) |
|---|---|---|---|---|---|
| CodeAF (senior-dev) | 21 | 20 | 15 | 3 | 3 |
| mini-swe-agent | 17 | 18 | 16 | 3 | 2 |
| Codex CLI | 17 | 17 | 13 | 1 | 3 |
| Pi | 12 | 15 | 12 | 3 | 0 |
| Oh My Pi | 5 | 11 | 13 | 2 | 0 |
| OpenCode | 7 | 10 | 10 | 3 | 0 |
| Kilo Code | 7 | 12 | 10 | 1 | 0 |
| DeepSeek Harness | 5 | 5 | 6 | 0 | 0 |
| Claude Code | 0 | 8 | 6 | 2 | 0 |
| Muse Code | 1 | 2 | 0 | 0 | 0 |
Counts are tasks the official DeepSWE verifier passed, grouped by the language column of the per-task records. Bold marks the most in each column, including ties.
Method
The rows come from the comparison on the main DeepSWE results page: 113 tasks, the official verifiers at 0b9fabb, 3 hours per task, one pass per harness on one model through OpenRouter. This page groups the per-task records by language. The CodeAF rows are stored under senior-dev.
Go
CodeAF solved 21 of the 34 Go tasks (62%). mini-swe-agent and Codex CLI solved 17 each (50%), and Pi 12. That four-task lead is the largest CodeAF had in any language. Claude Code, driving DeepSeek V4 Flash rather than a Claude model, solved none of the Go tasks.
Python
CodeAF solved 20 of the 34 Python tasks (59%), against 18 for mini-swe-agent and 17 for Codex CLI. A two-task gap on one pass is close; read it as "among the leaders" rather than a settled first place.
Where the other tools did better
- TypeScript. mini-swe-agent solved 16 of 35 TypeScript tasks and CodeAF 15, so CodeAF was second on the largest language group.
- JavaScript and Rust are ties. With five tasks each, CodeAF tied four harnesses on JavaScript (3) and Codex CLI on Rust (3). Five tasks are too few to rank anyone.
Limits of this result
- It is AgentField's comparison, one pass each, not a public leaderboard. A gap of two or three tasks inside one language is within run-to-run noise.
- CodeAF used the provider's default sampling; the other nine sent temperature 1.0 and top-p 0.95. CodeAF ran a day later, on 2026-09-12 UTC.
- The language comes from the task records. Some tasks touch more than one language; each is counted once, under its recorded language.
- Results are for DeepSeek V4 Flash. A stronger model lifts every harness; this ranks the harnesses on one model.
Questions people ask
What is the best AI coding agent for Go?
In AgentField's DeepSWE comparison, with ten harnesses on DeepSeek V4 Flash, CodeAF in senior-dev mode solved the most Go tasks: 21 of 34, against 17 for mini-swe-agent and Codex CLI. It is one benchmark and one pass, so weigh it with your own trial on your codebase.
What is the best AI coding agent for Python?
On DeepSWE's 34 Python tasks with every harness on DeepSeek V4 Flash, CodeAF solved 20, mini-swe-agent 18 and Codex CLI 17. The lead is two tasks on one pass, so CodeAF, mini-swe-agent and Codex CLI are best read as the leading group.
Which coding agent is best for TypeScript?
On DeepSWE's 35 TypeScript tasks, mini-swe-agent solved 16 and CodeAF 15, with Oh My Pi and Codex CLI at 13. On this benchmark and model, mini-swe-agent was narrowly first.
Can CodeAF use a different model for Go or Python work?
Yes. CodeAF works with the model you choose, including OpenRouter, DeepSeek, Kimi, Qwen, a Codex plan or a local model through Ollama. Any model explains how to pick one.
Sources
- Per-task records, checked 2026-11-21.
- DeepSWE, for the task set, checked 2026-11-21.
Last updated 2026-11-21. Every count recomputed from the per-task CSVs.