Coding-agent gains now come from engineered context and executable checks

Repository understanding is becoming an explicit stage with its own retrieval and cost choices. ACQUIRE asks targeted questions before patch generation and raises Pass@1 on SWE-bench Verified by 3.8 to 4.4 percentage points over Mini-SWE-Agent. A separate localization study finds that role-aware file summaries improve Hit@5 by up to 40% over paths while using 10.4 to 20.9 times less text than raw code. FlowArk applies the same selectivity across sessions: it injects verified knowledge only when code anchors match, reducing Android analysis cost by 26.83% at comparable quality.

Agent reliability work is putting checks around tool output and generated changes. TerraRepair consults Terraform dependencies and installed provider schemas, then reruns a security scanner before accepting a patch. Its Checkov-verified AWS fix rate reaches 78.4%, compared with 26.6% for the controlled one-shot baseline, though missing deployment context still triggers escalation. AgentCheck complements this with replayable tool faults: agents passed 77 to 105 of 120 scenarios, and failures usually involved confident use of bad output rather than crashes. The need for these controls is visible in a 44-developer study where Copilot improved functional correctness but did not significantly improve secure API usage.

Work on agent memory continues to favor deterministic, inspectable state. Context Warp Drive folds old turns without another model call, preserves exact identifiers, and reports a 92.6% cache-read rate in one production deployment; its real-workload evidence lacks a controlled comparison. Daftari stores memory as Git-backed Markdown with provenance, supersession links, and unresolved contradictions, but provides no formal benchmark. FlowArk offers stronger task-level evidence for selective reuse: bounded knowledge tied to code matches completed 1,060 Android analysis tasks under a $100 budget, versus 776 without reuse.