AI Code Ships 78% More Prod Incidents Than Human Code

Behind the 78% number: AI code is optimized to pass review, not survive production. It compiles, reads clean, and matches idiom, so reviewers rate it higher than human code. Meanwhile it reproduces vulnerable training-data patterns and hallucinates dependencies. Veracode pegs that at roughly 50% insecure output. The feedback loop inverted. The cheap early signal, review, lies. The expensive late signal, an incident, is the only honest one.

Where sources converge

New Relic, Veracode, and GitHub each point the same way. New Relic quantifies the incident gap. Veracode's dependency-hallucination vector is the one that keeps me up: a hallucinated package creates a namespace a malicious actor can register, and now it's a supply-chain hole. And GitHub's Copilot review agent got worse with more tools. Generic instructions made it browse the whole repo instead of the diff. Rewriting the prompt to "start from the diff, narrow with grep/glob, read exact evidence" cut cost 20% with no quality loss. Under-specified AI explores exhaustively and confidently reports noise.

The bifurcation

Human review keeps its value for design, mentorship, and architecture. It fails as bug-catching at AI-PR velocity. Split the roles. Humans review intent and structure. Machines enforce correctness via property-based tests, contract tests, and mandatory SAST/SCA that cannot be bypassed by human approval.

At AI-generated PR velocity, code review as a bug-catching mechanism is theater. The real gate runs in the pipeline.

Concretely: instrument AI-authored and human-authored code paths separately, then correlate incident rates. Verify resolved packages against known-good hashes so a hallucinated dependency can't resolve. Treat AI output as untrusted until validated.

All three findings here came with production evidence behind them.

DuckDB at 170TB is not a demo

Arcesium moved a 170TB, 15-trillion-record P&L warehouse off a traditional RDBMS onto Iceberg-on-S3 with DuckDB compute. They open-sourced the execution layer, Swiftlake. The usual objection is that DuckDB is per-process. That stops mattering when the workers scale horizontally via KEDA on Kubernetes, one partition per worker. Iceberg time-travel takes over the recovery job the RDBMS transaction log used to do. Ingestion came in 80% faster, egress timeouts went away, cost dropped 40%. If you're paying for always-on RDBMS compute on large analytical tables, this is worth a spike.

pgrust: audacious, but read the trade-off

pgrust boots from existing Postgres data directories and passes 46,000+ Postgres 18.3 regression queries. It swaps process-per-connection for threads, adds 64-bit TXIDs that kill wraparound freezes, and ships an adaptive planner. The 50% faster transactions claim is plausible. The 300x on analytics needs extraordinary evidence. Here is the cost of the thread model: process isolation is a safety feature. In stock Postgres a corrupted connection can't take down its peers. Threads give that up. Treat it as "assess." Run its regression suite against your schema. Don't plan a migration for 2-3 years.

Ingress toil, quantified

A logistics team on EKS, 12 microservices at 35K–61K RPM, picked Traefik over ingress-nginx even at 11% more memory. Route config went from 23 minutes to 4, worth roughly 6 operator-hours a month. The failure mode showed up at cutover: a misconfigured timeout annotation drove an 8% error spike. The fix is an annotation-mapping doc and production traffic replay in staging before you flip it.

DuckDB-plus-Iceberg proves the horizontal-worker pattern beats building a distributed query engine at 170TB scale.

Agent tooling moved from capability demos to infrastructure economics this week. The interesting parts now come with numbers.

Fix trace topology before it bankrupts you

Prime Intellect's verifiers v1 stores rollout traces as message DAGs, not linear histories. Here's what the naive path actually does: it re-stores the full conversation each step. That's O(n²) in turn count. DAGs store each message once and reference by pointer, which is O(n). A multi-turn agent with unexamined trace storage is a cost bomb that compounds quietly until the bill arrives.

Instructions as a funnel, not a toolbox

GitHub found that adding more tools made review worse, which is counterintuitive until you trace it: broad capability plus vague instructions means the agent explores exhaustively, burns tokens, and generates noise. The fix was "start with the narrowest context, expand only on specific evidence." That cut Copilot review cost 20% with no quality loss. Glean's code-based orchestration — the LLM generates a program that runs deterministically instead of reasoning between each tool call — cut tokens 24%. It's query planning for agents: compile the plan once, then execute it cheaply.

Parallelize the mechanical work

Sierra's git-worktree pattern gives each agent its own working directory on the same repo, and reports 5x throughput on genuinely independent tasks like test backfill, mechanical refactors, and boilerplate. The constraint is task decomposition, not tooling. Cross-cutting work creates merge hell. Expect 2-3x sustained, not 5x.

Intelligence is commoditizing; your orchestration topology and captured judgment are the moat.

Microsoft's tool-boundary guardrails and the memory-sidecar research point the same direction. Decompose agents into specialized, testable units with clear interfaces. Cap every autonomous loop with an explicit budget. Capture correction signals as portable eval data, because that's the part a competitor can't lift from your spec sheet.