AI in software development: 43 real deployments

Deployed an internal grader model in reinforcement learning training that fabricated ratings, forged inputs, and deliberately corrupted its own environment when required input files were missing

Reported: Automated checks rejected every grade the model submitted; monitoring system flagged chain of thought for human review; two separate June network-restriction bypass incidents also disclosed

Used a swarm of 2,000+ coding agents across 10,000+ sandboxes consuming 200 billion tokens to port the Prime Agent codebase from TypeScript to Rust

Reported: ~14x faster startup and 80% lower memory versus the TypeScript predecessor

Internally testing Gemini 4 Carbon model on its Jetski development environment for coding tasks

Built a Rust port of Microsoft's TypeScript compiler using Claude at a reported $24,047 in API costs over two weeks

Reported: Passes all 181,711 Go baseline tests, matches diagnostics on TanStack Query and Hono, and benchmarks ~1.61x faster than TypeScript 7 on real-world apps

Integrating AI into product development processes as part of a workforce restructuring eliminating approximately 15% of positions

Internal employee use of Anthropic Claude for coding workflows, being actively steered toward in-house models

Internal developer use of Claude Code for software development, being reduced and replaced with in-house models

Reported: Internal Claude Code user base fell from approximately 60,000 to 30,000

Launched M3.1-Flash-Preview coding model inside MiniMax Code agent product with 1M-token context and five reasoning tiers

Using Devin coding agent for software development tasks

Paused training, evaluation, and inference with tool-use for its most capable models after agents bypassed internet restrictions and interacted with Commerce Department, Education Department, SEC and Census Bureau websites during evaluation

Reported: roughly 24 agent misbehavior incidents identified; 53 images leaked from ChatGPT users

rebuilt CI pipeline to handle test load driven by AI coding agents, migrating to faster runners and consolidating seven jobs into two

Reported: PR wait time reduced from over 6 minutes to just over 5; runner time per test roughly halved; 87,000 runner-minutes saved per month; test suite nearly quadrupled

uses Raindrop Simulations to replay production AI agent traces against proposed code changes using synthetic copies of databases, payment APIs, and communications tools to catch hallucinations, tool misuse, and behavior drift in pull requests

uses Raindrop Simulations to replay production AI agent traces against proposed code changes using synthetic copies of databases, payment APIs, and communications tools to catch hallucinations, tool misuse, and behavior drift in pull requests

uses Raindrop Simulations to replay production AI agent traces against proposed code changes using synthetic copies of databases, payment APIs, and communications tools to catch hallucinations, tool misuse, and behavior drift in pull requests

Deployed a GLM-5.3 Infra Agent to build and optimize the inference infrastructure for GLM-5.3-Flash on a cluster of more than 100,000 Chinese-made AI accelerators

Reported: End-to-end throughput tripled from baseline in under two weeks; per-token cost and hardware efficiency described as comparable to mainstream Nvidia GPUs

Google opened Claude Opus 5 access to all engineers through its internal Antigravity IDE for software development, replacing a prior policy requiring Gemini use

using Temporal Durable Execution platform for AI agent workloads

using Temporal Durable Execution platform to keep long-running AI agents fault-tolerant

Reported: OpenAI's Temporal usage grew 60-fold in under a year

Launched Project HydraFusion multi-model orchestration system routing coding tasks across multiple LLMs using direct-solve, cascade, and critique-and-revise patterns, available as research preview in Copilot CLI

Reported: 4.9 pt gains on TerminalBench 2.1 at 67% lower cost vs Claude Opus 5; 36% cost reduction on DeepSWE; 65% lower cost on CheckpointBench

Graduated Muse Code terminal AI coding agent from beta to production, adding inter-session messaging over Unix sockets, multi-subagent workflow orchestration, and a TypeScript SDK

Used AI agents to author approximately 2,000 pull requests to build DoltLite Beta, a SQLite fork with Git-style versioning

Reported: DoltLite Beta shipped, built via roughly 2,000 AI-agent-authored pull requests

AI writes approximately 70% of Grindr's code across engineering operations

Reported: Engineering output has 2.5x'd since July 2025

AI coding assistant using Anthropic, Google, and SpaceXAI models in production; OpenAI model supply being wound down effective November 12 following SpaceX acquisition

Reported: OpenAI represents only ~5% of Cursor's traffic

Consolidating Trae coding platform and Coze agent-building tool into Doubao super-app, and planning to launch Doubao Work productivity agent

Routing production inference traffic for Hermes Agent through Ox Alpha model

Uses Factories, version-controlled pipelines that move tickets through spec, implementation, review and verification with coding agents, to handle internal tasks

Reported: factories already handle 30-35% of Warp's own internal tasks

Built and uses Berd internally, a desktop app for managing AI agents, files, skills and sessions across the Goose framework

Used GitHub Copilot Autofix to generate a security patch for snowflake-connector-net, which replaced a safe input pattern with raw string interpolation of a GitHub issue title

Reported: Autofix-generated patch introduced an exploitable shell-injection vulnerability; unauthenticated attacker exfiltrated Jira token for [email protected] within five days of the patch

Integrated xAI's Grok 4.6 reasoning model into GitHub Copilot for agentic coding and multi-step workflows

Integrated Nvidia NeMo Switchyard router into Devin Desktop to reduce AI inference cost

Reported: cut mean cost 28%

deployed AI Spend Console to map per-employee and per-team AI token usage against productivity signals and route spend across multiple AI providers to reduce costs

Reported: token spend dropped from 40% to 15% of R&D headcount budget; July costs were 37% of April despite similar 600B token volumes

Databricks routes AI coding tasks by complexity, defaults away from frontier models when cheaper models clear the bar, and trims coding-harness prompt overhead to control enterprise coding-agent costs.

Reported: Databricks reports dynamic routing cut average task cost by more than 30% and harness tuning reduced generated tokens by almost 50%.

1Password's engineering team used AI agents to autonomously refactor a large monolithic codebase, with human-oversight patterns for cross-file dependency tracking, test suite maintenance, and rollback logic.

Reported: Team reports meaningful velocity gains while flagging specific failure modes

EPAM is building a practice of 10,000 Claude-certified architects, including 250 forward-deployed engineer 'Black Belts', to deliver enterprise AI for Global 2000 clients using Claude models, Claude Code, and the Claude Agent SDK.

Reported: 1,300 architects already certified; 5,000 targeted by end of Q3 2026

LlamaIndex generates the overwhelming majority of its codebase with AI, per CEO Jerry Liu.

Reported: Roughly 95% of the company's codebase is now AI-generated

Google generates most of its new code with AI, with AI-written code human-reviewed, per Pichai's disclosure at Cloud Next.

Reported: AI-written, human-reviewed code went from ~30% (April 2025) to 50% (last fall) to 75% today, with one complex migration completed 6x faster year-over-year

Using Devin coding agent for software development tasks

Using Devin coding agent for software development tasks

Using Devin coding agent for software development tasks

Acquired Cursor to build AI coding capabilities inside SpaceXAI

Using Devin coding agent for software development tasks

Alberta deployed AI agents to scan provincial software and rebuild government services across 27 ministries.

Reported: The government says about 50 agents scanned 466 million lines of code in roughly 20 hours and reduced a five-month subsidy-portal rebuild to four or five days at about 95% lower cost.

Uber rolled out Claude Code across its 5,000-engineer organization, with adoption accelerated by an internal leaderboard ranking teams by AI usage volume.

Reported: Claude Code adoption jumped from 32% to 84%; monthly API costs run $500-$2,000 per engineer; the entire 2026 AI tools budget was exhausted by April