Slack Code, Agent Harnesses, Review Bottlenecks
Industry
Dev Tools
Overview
AI coding coverage is moving from code generation to the systems that make agent output usable: shared workflows, harnesses, review capacity, identity, and data controls. Slack Code's launch shows agents moving into team collaboration, while NVIDIA's AVO research suggests the surrounding harness can matter as much as the model. The countertrend is operational: trust in AI accuracy fell to 29% in cited Stack Overflow data, the best result on a new refactoring benchmark was 41.2%, and security incidents exposed risks at build and sandbox layers.
Key Stories
Shared workspaces are becoming the next interface for AI coding. InfoWorld, VentureBeat, and theregister covered Slack Code, which places Claude Code, Devin, GitHub Copilot, ChatGPT, and Vercel's agent in shared channels for planning, review, previews, and pull requests. The model could reduce handoffs and make agent work more visible, but coverage questions whether Slack is suitable for deep engineering work and warns of interruption and channel-sprawl costs. "Coding is deep work, and Slack is the interruption machine," said a Broadcom engineer.
The agent harness is emerging as a distinct product and cost layer. TechCrunch and NVIDIA Technical Blog described AVO, which combines memory, tools, feedback, supervision, and execution loops. NVIDIA reported a 30% to 100% improvement for Claude Opus 5 on the cited ARC-AGI-3 test, while AVO also completed 183 levels after seven days of GPU-kernel work. The implication is that runtimes, evaluation, observability, and recovery may differentiate developer tools even when models are interchangeable. "That itself can 2x your cost," Databricks CEO Ali Ghodsi said of choosing the wrong harness.
Review and verification, not generation, are becoming the adoption bottleneck. Latest news, DevOps.com, and The New Stack linked AI coding adoption to workload, trust, and code-quality concerns. Cited data shows 80% of developers use AI tools, but trust in accuracy has fallen to 29%; SWE-Bench ProMax's best model resolved only 41.2% of 170 large-scale refactoring tasks. Tessl's standards-as-code approach and the growing cleanup market point to opportunities for review, testing, refactoring, and policy tooling around agent output.
Security controls are shifting outside the model. The New Stack reported that a malicious pull request reached nearly one million Visual Studio Code marketplace users, while separate coverage described agents probing 9,000 hosts and compromising four companies through sandbox or supply-chain weaknesses. SD Times highlighted AI SAST that reportedly found 96 of 102 known C vulnerabilities. The common requirement is independent enforcement through scoped credentials, external policy gates, signed releases, runtime identity, and human approval.
Free and anonymous models could increase experimentation, but not confidence. TechCrunch and Business Insider covered Ox Alpha, a temporarily free model attracting developers for coding and long-running agent tasks. Its origin is unconfirmed, with speculation involving Z.ai, Xiaomi, and Microsoft. The development may expand low-cost experimentation, but provenance, reliability, and governance remain unresolved.
