The New York Report — three days at AI Engineer New York
Field report from AI Engineer New York (October 12–14, 2026): why 50% of large enterprises run coding agents while only 5% have autonomous agents in non-engineering operations—and the bespoke evals, policy DAGs, bitemporal ontologies, VM sandboxes, and KV-cache tokenomics crossing that gap. Published as STATE No. 02.
Three days of sessions in Midtown Manhattan at AI Engineer New York (October 12–14, 2026), paired with our complete 31-talk interactive archive. Across Fortune 500 enterprises, 90% have rolled out chat LLMs and 50% run coding agents, yet roughly 5% have autonomous agents working in non-engineering operations—while inside engineering itself, a 2,450% surge in AI-written code yielded only 30% more releases because human reviewers accept wrong AI output 80% of the time.
The four engineering shifts crossing that 95% diffusion wall:
- Bootstrapping domain evals and verification without public benchmarks. Teams invert dataset creation around known source pages and synthetic contract breach clauses (Air Canada), backtest web signals over 400B+ dated page snapshots without lookahead bias (Exa), grade multi-hour workflows with ground-truth-free behavior specs (Basis), invert author/reviewer model families for +10% bug detection (Greptile), and gate completion behind external MCP verifiers paired with System One classifiers (WorkOS).
- Bounded autonomy, policy DAGs, and governed financial workbenches. Regulated firms replace air-gapped sandboxes and prompt wishes with automatic framework registries and compiled runtime tool entitlements (Two Sigma), parallel compliance policy DAGs (Coinbase), 100M-document bitemporal ontologies built for 1,000-QPS agent storms (
Vaultat Bridgewater), and golden-source reconciliation across 20 conflicting internal systems (Apollo + Rowspace). - Isolated Linux VMs, outer-loop institutional context, and the shrinking harness. Frameworks are deleting local host executors in favor of
create / execute / deleteVM sandboxes (Daytona) and self-deleting guardrail shims (SpaceX), while curing product incoherence by threading customer intent and historical reverts through automated triage, review, living wikis, and $1,000 legacy war rooms (Linear, Greptile, GEICO). - Agent managers, trace-to-skill loops, and real tokenomics. Engineers move off
localhostto manage cloud builder/QA/merge-manager fleets gated on feature-demo videos (Untangle) and post-mergeSKILL.mdrubric PRs (Warp), while slashing inference spend 10× to 20× via per-department API keys, prefix-KV-preserving task routing (2× speed ≈ 3× cost), 20–30× offline KV-cache cartridges, and post-trained 35B open specialists (Featherless, Ramp, Runpod, Nous Research).
The complete six-chapter monograph is bound as STATE № 02 · The New York Report, alongside the searchable 31-talk field archive and 8 theme atlases.