Executable proof is becoming the standard output for coding agents

Execution is now the gate for agent claims. AgentForge requires every patch to run in a network-isolated Docker sandbox before it can move forward, and reports 40.0% resolution on SWE-bench Lite, with a 26 to 28 point gain over its single-agent baselines. AnalysisBench reaches the same conclusion in a different setting: agents need explicit stages and evidence-based stopping rules, because self-validation still overstated success by 15%. AnyPoC applies the pattern to security reports by generating and re-running executable proof-of-concept tests, then rejecting bug reports that fail that check.

The strongest systems are writing down workflow boundaries, not just adding more model capacity. AnalysisAgent uses one action per cycle, deterministic log condensation, and manual verification of tool-specific outputs across 35 tool-project tasks. CodeTracer studies what happens when those boundaries are missing or weak. It reconstructs runs into state-transition trees, then locates the earliest failure-causing stage and step. On GPT-5 traces, step-level F1 rises to 48.02 from about 19 for lighter baselines, while token use drops to 31.1k from 44.8k to 58.5k. The practical message is clear: explicit stages help both execution and diagnosis.

This period also extends executable checking beyond coding benchmarks into harder real settings. AnyPoC works across 12 large software systems, including Chromium, Firefox, LLVM, OpenSSL, SQLite, FFmpeg, and Redis. It reports 1.3x more valid PoCs for true bug reports, 9.8x more rejected false positives, and 45 generated PoCs adopted as official regression tests. ORBIT, in a separate translation setting, couples project-level orchestration with compile and test repair, reporting 100% compilation success and 91.7% test success on 24 CRUST-Bench programs. The common thread is operational proof: compilation, tests, traces, and reproducing artifacts are becoming the accepted output, not just plausible text.