Agent products advertise controls faster than public evidence

Minovative Mind CLI is framed as an agent that can edit a repository without leaving it in a broken state. Its claimed controls include local semantic code search, dependency tracing across 11 languages, parallel sub-agent work with a mutex registry, pre-flight syntax checks, fuzzy patching, transaction logs, sandboxed build trials, and /revert rollback. The build loop is bounded: up to 120 seconds per trial and up to 5 auto-correction attempts after compiler errors or regressions.

The evidence is feature-level. The source gives no SWE-bench, HumanEval, RepoBench, pass rate, latency, cost, or baseline comparison. That makes it useful as a map of current coding-agent product requirements, not as proof that the system outperforms other agents.

Aikido Code Audit targets vulnerabilities that require following intent and state across files. The product scans static source across one or more repositories, follows references across modules, and returns findings with root cause, code evidence, and an AutoFix pull request. The examples include a multi-file insecure direct object reference chain, source-level ReDoS detection, and admin-only routes that a live test may miss.

The claims are specific but mostly vendor-reported. Aikido says early use found a median of about 25 issues per codebase, that no audits came back clean, and that the tool covers roughly 70–80% of what a full pentest finds at about one-tenth the cost. The excerpt does not provide a public dataset, reproducible protocol, or independent evaluation.

StaleTrace focuses on a practical failure mode for deployed agents: acting on a fact after that fact changed. It ingests tool calls and recorded fact events, replays them into a temporal ledger, assigns validity windows, and checks what the agent used against what was valid at that moment. The output is a root cause, blast radius, and incident report.

The design choice is determinism. StaleTrace claims no large language model calls, no embeddings, and no graph database during auditing. The same inputs are claimed to produce the same verdict. The source shows one reconstructed incident example and gives no accuracy, latency, dataset, or production-volume measurements.

The model-routing article argues that many office tasks can be handled by small, domain-tuned language models selected by a cheap router. Its reported GDPVal-AA setup sends hard tasks to GPT-5.5 and easier tasks to GPT-5.4 Mini. The routed system scores 1759 ELO, close to GPT-5.5 alone at 1769, while GPT-5.4 Mini alone scores 1417. The router overhead is reported as less than $0.01 per request, and the selected model is locked for the session to preserve prompt caches and output consistency.

The article pairs routing with domain post-training examples. It reports that MAI-Code-1-Flash, with about 5B active parameters, scores 51.2% on SWE-Bench Pro versus 35.2% for Claude Haiku 4.5 and uses up to 60% fewer tokens. It also claims the routing plus tuned small-language-model setup can cut cost by 75–90% and improve latency by 2–3×. These numbers make cost-quality tradeoffs the main evaluation claim.