Home About Spatial Lab Disciplines Agentic Tools
Learn AI
IP Network Infrastructure Blog Connect
Autonomous Software Engineering Lab SWE-bench Verified MicroVM Sandboxing Pass@k Evaluation

Autonomous Coding Agent & SWE-Bench Execution Lab

Model real-world repository ingestion, AST symbol retrieval, autonomous git diff generation, test execution feedback loops, and evaluate pass@k resolve rates across frontier coding reasoning engines.

Agent Configuration SWE-bench Runner

Repository Context Budget 64,000 tokens
16k 96k 192k tokens
Self-Healing Iterations (Max Turns) 4 Iterations
1 (Single-Shot) 4 (Standard) 8 (Deep Repair)
Candidate Samples (k for Pass@k) k = 5 samples
k=1 k=5 k=10
Projected Pass@k Rate
88.4%
k=5 with self-healing
Run Cost (API & Sandboxing)
$0.42
128k input + 6.2k output
Sandbox Provisioning Latency
5.2 ms
Hardware KVM Isolation
Total Wall-Clock Runtime
18.4 s
3 LLM turns + 2 test executions
SWE-Bench Execution Harness Lifecycle RESOLVED • ALL TESTS PASS
STAGE 01
AST Ingestion
0.8s • Tree-Sitter
STAGE 02
Patch Synthesis
6.2s • Extended CoT
STAGE 03
MicroVM Sandbox
4.1s • Pytest Harness
STAGE 04
Self-Healing Repair
7.3s • 1 Feedback Loop
--- a/django/core/handlers/base.py +++ b/django/core/handlers/base.py @@ -124,7 +124,9 @@ class BaseHandler: for middleware_method in self._view_middleware: - response = middleware_method(request, callback, callback_args, callback_kwargs) + if asyncio.iscoroutinefunction(middleware_method): + response = await middleware_method(request, callback, callback_args, callback_kwargs) + else: + response = middleware_method(request, callback, callback_args, callback_kwargs) if response: break
Reasoning Engine SWE-bench Verified (Pass@1) Pass@5 (Self-Healing) Avg Turns to Resolve Cost Per Solved Issue
Claude 3.7 Sonnet 70.3% 88.4% 2.8 turns $0.58
OpenAI o1 68.4% 86.1% 3.1 turns $1.45
OpenAI o3-mini (High) 64.2% 82.0% 3.4 turns $0.24
DeepSeek-R1 62.8% 80.5% 3.6 turns $0.18

Autonomous Software Engineering Curriculum