BreakpointTL;DR: Now that nearly all of our new code is AI generated and overseen by humans, we’ve rebuilt our engineering interview loop to keep pace, testing how candidates direct AI, evaluate its output, and apply judgment where models fall short. This post documents three phases of that work: what changed, what the data showed, and what we're still figuring out.
Coinbase engineering blog
TL;DR: This post dives into how Coinbase’s Risk ML team uses Agentic Skills to close the model improvement loop with human review gates.
TL;DR: When a new fraud pattern hits, figuring out which signals already exist in your data — and what new ones you should build — is a manual, time-consuming process. We built a search tool over our internal feature store and event logs that lets fraud analysts describe a new scam in plain English and quickly discover relevant signals.
TL;DR: On May 7, 2026, an AWS thermal event triggered a Coinbase outage. Recovery was delayed by a matching engine locked to the failed zone and a silent failure in AWS's managed Kafka service. Moving forward, we are improving our cross-zone standbys for our low-latency exchange and updating Kafka infrastructure.
This report summarizes Coinbase’s Ethereum validator performance for Q1 2026, highlighting strong uptime, zero slashing incidents, and broad operational diversity across clients, relays, regions, and cloud providers. It also explains Coinbase’s staking philosophy and the operational safeguards it uses to support network resilience, decentralization, and institutional-grade staking performance.
TL;DR: AI coding agents made individual engineers faster, but the workflow around them stayed sequential. That’s changed. At Coinbase, engineers are evolving from implementers into orchestrators of agent fleets. The evidence is Mux, an internal multi-agent tool that started as one engineer's side project and grew organically to 600+ users across every org. Power users now merge 3.5x more PRs than baseline.
TL;DR: At Coinbase, ML models handle long-term fraud defense while rules let us respond fast to active attacks. We rebuilt the backtesting data layer, automated schema evolution, and gave analysts a standardized notebook workflow backed by ML libraries. Backtesting is now faster, and analysts can go from spotting a fraud pattern to shipping a new rule in hours, not days.
This post explains how Coinbase scaled Ethereum staking exchange-traded products by translating validator-level protocol behavior into portfolio-level financial operations. It covers the operational impact of EIP-7251, off-chain vault transfers to solve liquidity constraints, intent-based portfolio staking, and deterministic reporting across consensus and execution layers for daily NAV calculations.
Coinbase built an AI-powered Case Grading Assistant to cut trainee case review time from about 90 minutes to about 20 minutes. The system uses comparison-based grading against reference cases to improve consistency while keeping human oversight and auditability.
TL;DR: Coinbase processes over a billion payout transactions a year across staking rewards, USDC rewards, and Coinbase One benefits. We’re sharing the evolution of our Payout Framework—from the limitations of synchronous processing to a high-throughput async architecture that handles rewards across dozens of assets with precision.
This post explains Distributed Proofs of Possession (DPoP), a cryptographic method Coinbase uses to prove control of key shares without ever reconstructing the full private key. It outlines why using production MPC signing for audits is risky, then shows how independent share signatures and public-key reconstruction provide sound, zero-knowledge verification while remaining unusable as transaction signatures.
**Tl;dr**: Coinbase has developed a novel, dynamic control policy that replaces static rules to automatically manage risk, resulting in superior financial loss mitigation and more efficient utilization of constrained resources.
Coinbase engineered MagicSpend and Magic Platform to simplify the fragmented infrastructure needed for retail onchain products. Together, these layers reduced shipping timelines and processed more than $1B in volume.
TL;DR: By leveraging a "Context-First" AI workflow, we shipped Coinbase Business Invoicing in weeks instead of months. We transitioned from manual UI mocking to code-first prototyping and used scoped repository sandboxing to enable safe, cross-stack iteration.
**Tl;dr:** To meet the scaling demands of Solana, Coinbase has moved away from its legacy chain-agnostic processing model. We engineered a dedicated, high-throughput streaming architecture with parallel block processing, resulting in a 12x increase in transaction processing throughput and a 20% reduction in deposit latency.
We built Smart Targeting to move beyond using manual segmentation to build target audiences towards automated, intelligent discovery.
This post explains how Coinbase scales personalization by replacing hand-crafted feature engineering with user foundation models trained on long-range sequences of user activity. It details a self-supervised Transformer-based approach, compares methods for applying embeddings to downstream tasks, and shows how the system improves both model performance and development speed while helping solve cold-start targeting problems.
This post explains how Coinbase builds enterprise AI agents with a focus on trust, scale, and repeatability. It highlights the engineering practices needed to make AI systems reliable in production, including strong safeguards, operational consistency, and workflows that can be reused across teams.
Coinbase announces a system update highlighting several new products launched to advance its vision for the future of finance. The post frames these releases as part of a broader effort to expand Coinbase's platform and services for users.
This post explains how Coinbase replaced a vendor-managed futures risk system with an in-house solution called STARK to improve consistency, latency, and operational resilience. It highlights the architectural shift to atomic snapshots, on-premise processing, Kafka-based recoverability, and extensive shadow-mode testing that enabled a zero-downtime migration and stronger support for 24x7 trading.