Andrew Ng: AI Coding Agents Demand Judgment Over Prompting

The new AI engineering hierarchy

Andrew Ng has published a map identifying the four most valuable AI engineering skills based on analysis of more than 10,000 job postings and interviews with AI experts, hiring managers, and recruiters. The findings, first reported by Forbes contributor Josipa Majic Predin, reveal a sharp turn away from technical prompting toward judgment-based capabilities.

Two of the four skills—using coding agents and shaping the build—describe work that had no hiring category in 2022. Neither requires writing prompts. Both center on orchestrating systems and making architectural decisions under uncertainty.

The timing matters. Global venture funding hit $510 billion in the first half of 2026, with more than 70 percent of second-quarter capital flowing to AI companies. OpenAI and Anthropic alone absorbed $217 billion, claiming 43 percent of all startup funding tracked by Crunchbase. Application-layer founders compete for the remainder in a market where capital is abundant but selection criteria have hardened.

Why it matters

As AI-assisted coding collapses prototyping costs, investors are shifting focus from technical depth to judgment and evaluation discipline. Teams that can orchestrate agents, shape builds strategically, and measure productivity gains accurately will capture disproportionate funding and talent. The skills that mattered in 2022 no longer predict which startups ship.

Meta reprices agent access to capture training data

Meta introduced Muse Code, a command-line agentic harness, alongside its Muse Spark 1.2 model. Standard access costs $1.25 per million input tokens. A contributor tier costs $0.10—a 92 percent discount—in exchange for Meta's right to train on prompts and outputs. The contributor tier caps at 100 requests per minute per team versus 3,000 on standard, confining it to individuals and small teams whose entire product often sits in the repository the agent reads.

On Vals AI's Finance Agent v2 benchmark, which assigns models entry-level financial analyst work, Muse Spark 1.2 at maximum reasoning ranked first among 45 models at $0.77 per task, ahead of Claude Opus 5 at $5.12. On the broader Vals Index it placed fifth at $0.70 per task, above Claude Opus 4.8 at $7.52.

Mark Zuckerberg explained the architecture advantage: subagents persist across sessions inside isolated worktrees and edit in parallel. The harness logs every model call and file edit locally so crashed runs resume instead of restarting.

Venture studios bet on collapsed build costs

Ng's AI Fund, which holds more than $370 million from Sequoia Capital, NEA, SoftBank, and Nikkei, co-founds companies at pre-seed rather than writing passive checks. After closing a $190 million second fund, Ng credited the model to prototyping costs collapsing under AI-assisted coding.

Measuring AI's actual productivity gains remains difficult. Early studies showed mixed results; later ones suggest significant time savings that remain unverified. The uncertainty creates risk for investors evaluating teams, particularly as reduced building costs from agent use could blur ownership and contribution.

Ng's OpenWorker offers a local, user-controlled alternative to Meta's data-harvesting model, positioning privacy as a competitive advantage for teams unwilling to trade code access for discounted inference.

Details were first reported by Josipa Majic Predin in Forbes.