Key Signals

• IBM Bob reveals the enterprise AI coding playbook: intelligent model routing, not bigger models, is the real differentiator. Neel Sundaresan — founding engineer of GitHub Copilot turned IBM executive — disclosed that IBM Bob now serves 80,000 internal developers and routes tasks automatically across Anthropic Claude, Mistral, IBM Granite, and proprietary fine-tuned models based on task complexity. His observation that developers default to frontier models for trivial tasks underscores a cost discipline problem the entire industry faces. The architectural bet is that model orchestration and routing intelligence will matter more than any single model's capability. [1]

• The "paradox of supervision" is emerging as a structural risk for agentic coding adoption. An Anthropic study found a 47% drop-off in debugging skills among developers heavily using AI coding agents, while a LinkedIn engineering director has asked his 50-person team to avoid AI for tasks requiring critical thinking. The core contradiction: effectively supervising AI-generated code requires the very skills that atrophy from over-relying on AI agents, creating a feedback loop that may undermine the long-term viability of fully agentic workflows. This tension is forcing teams to reconsider where the human-agent boundary should sit. [2]

• Token economics are creating a new form of vendor lock-in distinct from traditional software dependencies. Unlike fixed employee costs, agentic coding costs fluctuate unpredictably with model pricing, token consumption patterns, and provider subsidies that may be withdrawn. Claude Code outages have already left entire engineering teams at a standstill, and developers report needing 2-3x more tokens after model updates to achieve the same results. The financial and cognitive dependency on model providers represents an industry-wide risk that local models cannot yet absorb at scale. [2]

AI Coding News

• IBM Bob's architect argues that model routing — not model size — is the future of enterprise AI coding. In a wide-ranging interview, Neel Sundaresan detailed how IBM Bob automatically routes tasks across multiple models — including Anthropic Claude, Mistral open-source, IBM Granite, and custom fine-tuned models — without exposing model selection to users. The system is optimized for enterprise conditions most AI coding tools treat as edge cases: legacy codebases, strict compliance requirements, and hybrid environments. Sundaresan warns the next frontier is agents communicating in machine-native languages humans cannot read, where "if there are errors in those derivative languages, that error could explode." He reports a 45% productivity improvement claim from the 80,000-user internal deployment. [1]

• A widely-shared analysis frames agentic coding as a "trap" that inverts developer priorities and accelerates skill atrophy. The article argues that Spec Driven Development — where developers define requirements and agents handle implementation — fundamentally inverts the traditional priority stack from understanding→standards→conciseness→speed to speed-first output. It cites multiple studies showing measurable cognitive impacts within months of heavy agent use, and quotes OpenCode creator Dax explaining that "typing out code is the process by which I figure out what we should even be doing." The proposed alternative is not abandoning AI tools but demoting them to secondary roles: using agents for brainstorming and bounded tasks while maintaining active engagement during implementation. [2]