Key takeaways
- 5% of all merged pull requests now come from Cloudbot background agents
- PR cycle time reduced from ~150 hours to ~15 hours (10x improvement)
- Multi-model architecture — explicitly not Claude-only, uses 'all sorts of underlying models'
- Skills and MCPs connect to Datadog, Sentry, Amplitude, Snowflake
- 800 engineers generated 300-400 PRs in 30 minutes during company-wide speedrun
FAQ
What is Coinbase's internal coding agent?
Cloudbot — a multi-model Slack-native coding agent serving 1,000+ engineers at Coinbase. It produces 5% of all merged PRs and reduced PR cycle time from ~150 hours to ~15 hours.
How does Coinbase's coding agent work?
Engineers tag Cloudbot in Slack or trigger it from Linear tickets. It uses Skills and MCPs (Datadog, Sentry, Amplitude, Snowflake) for context, creates plans, writes code across multiple repos, and delivers PRs with Cursor deep links and QR codes for mobile testing.
What percentage of Coinbase PRs come from AI agents?
5% of all merged pull requests at Coinbase come from Cloudbot, with PR cycle time reduced 10x from ~150 hours to ~15 hours.
Is Coinbase's coding agent built on Claude?
No. Despite early confusion around the name, Cloudbot is explicitly multi-model — 'it's actually using all sorts of underlying models. It's not something that is specific to Claude.'
Executive Summary
Coinbase's Engineering VP Chintan Turakhia announced in February 2026 that 5% of all merged pull requests now come from internal background agents. The system, originally called "Claudebot," was built by just two engineers and uses a Slack-native interface where engineers tag the agent in any thread to invoke coding tasks. This represents significant production adoption at a major crypto exchange.
| Attribute | Value |
|---|---|
| Company | Coinbase |
| Type | Internal tool |
| Adoption | 5% of merged PRs |
| Team Size | 2 engineers (initial build) |
| Interface | Slack-native |
Product Overview
Coinbase's internal coding agent follows the emerging pattern of Slack-native background agents. Engineers tag the agent in any Slack thread, and it produces complete pull requests using the same tools and context available to human engineers. The system handles planning, debugging, and shipping end-to-end.
Key Capabilities
| Capability | Description |
|---|---|
| Slack-native invocation | Tag agent in any thread to start |
| Full workflow | Plans, debugs, and ships PR end-to-end |
| Same context as humans | Access to same tools and environment |
| Background execution | Works while developer focuses elsewhere |
Known Workflow
Slack thread (tag agent)
↓
Agent receives context from thread + links
↓
Planning and implementation
↓
Debugging and iteration
↓
Pull request ready for review
What We Know
Coinbase has disclosed limited technical details. What's publicly known:
| Aspect | Known | Unknown |
|---|---|---|
| Adoption rate | 5% of merged PRs | Total PR volume |
| Interface | Slack-native | Alternative interfaces |
| Team size | 2 engineers (initial) | Current team size |
| Origin | "Claudebot" | Current internal name |
| Tools | Same as human engineers | Specific integrations |
Public Announcement
From Chintan Turakhia (VP Engineering) on X, February 2026:
"5% of all merged PRs at Coinbase now come from our in-house background agents. Built last year by two engineers, the Slack-native framework gives agents the same tools and context as human engineers. Tag in any thread, agent plans, debugs, ships PR."
Strengths
- Proven adoption — 5% of all merged PRs is significant production impact at a major company
- Familiar UX — Slack-native interface meets developers where they already work
- Lightweight build — Two engineers built the initial version, showing it doesn't require massive investment
- Full workflow — Agent handles entire flow from task to PR, not just code generation
- Context integration — Same tools and context as human engineers reduces friction
Cautions
- Limited public detail — Architecture, validation approach, and tooling not disclosed
- Scale context unclear — 5% impact depends on Coinbase's total PR volume (not disclosed)
- Evolution unknown — "Claudebot" name suggests Anthropic dependency; current architecture unclear
- Crypto-specific patterns — May include domain-specific integrations not transferable
- Not for sale — Internal tooling only
Competitive Positioning
vs. Other In-House Agents
| System | Differentiation |
|---|---|
| Stripe Minions | Stripe has 1,000+ PRs/week; Coinbase measures percentage |
| Ramp Inspect | Ramp at 30% adoption; Coinbase at 5% |
| Abnormal AI | Abnormal at 13%; Coinbase at 5% |
Adoption Spectrum
| Company | Metric | Value |
|---|---|---|
| Abnormal AI | % of PRs | 13% |
| Ramp | % of PRs | 30% |
| Coinbase | % of PRs | 5% |
| Stripe | PRs/week | 1,000+ |
Ideal Customer Profile
This is internal tooling, not a product for sale. The pattern is worth noting if:
Relevant indicators:
- Slack-centric engineering culture
- Interest in lightweight agent implementation (2-engineer build)
- Want to start with percentage-based adoption metrics
- Crypto/fintech regulatory environment
Limited applicability:
- Need detailed architecture guidance (Stripe/Ramp better documented)
- Want to eliminate human review (consider StrongDM)
- Require specific technical specifications
Viability Assessment
| Factor | Assessment |
|---|---|
| Public Documentation | Limited (X announcement only) |
| Adoption Metrics | Clear (5% of PRs) |
| Architecture Detail | Not disclosed |
| Replicability | Unknown |
| External Validation | Limited |
The 5% metric provides a useful benchmark for organizations measuring background agent adoption, but the lack of technical detail limits deeper analysis.
Bottom Line
Coinbase's 5% PR adoption demonstrates that background coding agents can achieve meaningful production impact with relatively small initial investment (two engineers). The Slack-native pattern matches what we see at Stripe, Ramp, and Spotify.
Key metric: 5% of merged PRs from background agents.
Key insight: Lightweight build (2 engineers) can achieve production adoption.
Recommended reference for: Organizations wanting adoption benchmarks, teams evaluating Slack-native patterns.
Not recommended for: Teams seeking detailed architecture guidance (see Stripe, Ramp instead).
Outlook: As Coinbase matures their system, expect more detailed documentation. The 5% benchmark provides a useful floor for what background agents can achieve at enterprise scale.
Research by Ry Walker Research • methodology
Disclosure: Author is CEO of Tembo, which offers agent orchestration as an alternative to building in-house.