Key takeaways
- 13% of pull requests now come from background agents — highest reported percentage
- Serves multiple personas: full-stack engineers, security analysts, and MLEs
- The agents/dev tools build themselves — self-improving infrastructure
FAQ
What percentage of Abnormal AI's PRs come from coding agents?
13% of merged pull requests at Abnormal AI come from internal background agents — the highest publicly reported percentage as of February 2026.
Who uses Abnormal AI's coding agents?
Full-stack engineers for feature development, security analysts for infrastructure patches, and MLEs (machine learning engineers) for agent tooling.
What infrastructure does Abnormal AI use for their agents?
Currently migrating from GitHub Actions to Modal for execution, suggesting a shift toward the Modal-based sandbox pattern seen at other companies.
Executive Summary
Abnormal AI (cybersecurity company) reported in February 2026 that 13% of their pull requests now come from internal background agents — the highest percentage publicly disclosed by any company. The system serves multiple personas: full-stack engineers building features, security analysts shipping infrastructure patches, and MLEs developing agent tooling. Notably, the agents/dev tools build themselves.
| Attribute | Value |
|---|---|
| Company | Abnormal AI |
| Industry | Cybersecurity |
| Adoption | 13% of merged PRs |
| Personas Served | Engineers, Security Analysts, MLEs |
| Infrastructure | Migrating from GHA to Modal |
Product Overview
Abnormal AI's background agents represent the highest publicly reported percentage of PR adoption. The system is notable for serving multiple distinct personas within the organization, not just software engineers. The self-improving aspect — where agents build their own tooling — suggests a maturing internal platform.
Key Capabilities
| Capability | Description |
|---|---|
| Multi-persona support | Full-stack engineers, security analysts, MLEs |
| Self-improvement | Agents build their own dev tools |
| Full-stack features | Complete feature development |
| Security patches | Infrastructure and security fixes |
| Agent tooling | Building the platform itself |
Personas and Use Cases
| Persona | Use Case | Example |
|---|---|---|
| Full-stack engineers | Feature development | New product capabilities |
| Security analysts | Infrastructure patches | Security fixes, compliance |
| MLEs | Agent tooling | Self-improving infrastructure |
What We Know
Abnormal AI has disclosed limited technical details. What's publicly known:
| Aspect | Known | Unknown |
|---|---|---|
| Adoption rate | 13% of merged PRs | Total PR volume |
| Personas | 3 distinct groups | Team sizes per persona |
| Infrastructure | Migrating GHA → Modal | Current architecture |
| Self-improvement | Agents build own tools | Specific mechanisms |
Public Announcement
From Shrividya Ravi (Head of Platform Engineering) on X, February 2026:
"13% of our PRs now come from background agents. We're migrating from GHA to Modal for execution. The agents serve full-stack engineers (features), security analysts (infra patches), and MLEs (agent tooling). The agent/dev tools build themselves."
Adoption Comparison
Abnormal AI's 13% represents the highest publicly disclosed percentage:
| Company | Metric | Value | Source |
|---|---|---|---|
| Abnormal AI | % of PRs | 13% | X announcement |
| Ramp | % of PRs | 30% | Blog post |
| Coinbase | % of PRs | 5% | X announcement |
| Stripe | PRs/week | 1,000+ | Blog post |
Note: Ramp's 30% may be frontend/backend specific, while Abnormal AI's 13% appears to be company-wide.
Strengths
- Highest reported adoption — 13% of PRs (vs. Coinbase 5%, Stripe volume-based) demonstrates significant production impact
- Broad use cases — Serves engineers, security analysts, and MLEs — not just developers
- Self-improving — Agents building their own tools suggests maturing platform
- Multi-persona design — Different user groups with different needs all served by same infrastructure
- Active evolution — Migration from GHA to Modal shows continuous improvement
Cautions
- Limited technical detail — Architecture not fully disclosed (blog post in progress)
- Infrastructure in flux — Currently migrating from GHA to Modal; patterns may change
- Cybersecurity domain — Patterns may include domain-specific elements not transferable
- No external validation — Limited coverage beyond X announcement
- Not for sale — Internal tooling only
Competitive Positioning
vs. Other In-House Agents
| System | Differentiation |
|---|---|
| Stripe Minions | Stripe measures volume (1,000+); Abnormal measures percentage |
| Ramp Inspect | Ramp at 30% (specific repos); Abnormal at 13% (company-wide) |
| Coinbase | Coinbase at 5%; Abnormal nearly 3x higher |
What Makes 13% Significant
- Company-wide metric — Not limited to specific repositories
- Multi-persona — Includes non-engineer users (security analysts)
- Self-improving — Platform builds itself, compounding gains
Ideal Customer Profile
This is internal tooling, not a product for sale. The pattern is worth noting if:
Relevant indicators:
- Multiple personas need coding assistance (not just engineers)
- Interest in self-improving agent infrastructure
- Cybersecurity or similar domain with compliance requirements
- Want highest-adoption benchmark to target
Limited applicability:
- Need detailed architecture guidance (wait for blog post or see Stripe/Ramp)
- Pure engineering organization (simpler patterns available)
- Require specific technical specifications
Viability Assessment
| Factor | Assessment |
|---|---|
| Public Documentation | Limited (X announcement, blog WIP) |
| Adoption Metrics | Strong (13% — highest reported) |
| Architecture Detail | Not disclosed |
| Multi-persona Design | Unique differentiator |
| External Validation | Limited |
The 13% metric and multi-persona approach provide useful benchmarks, but await the promised blog post for architectural details.
Bottom Line
Abnormal AI's 13% PR adoption represents the highest publicly reported percentage for background coding agents. The multi-persona approach — serving engineers, security analysts, and MLEs — suggests that coding agents can scale beyond pure software engineering use cases.
Key metric: 13% of merged PRs from background agents.
Key insight: Self-improving infrastructure where agents build their own tools compounds adoption gains.
Recommended reference for: Organizations targeting high adoption rates, companies with multiple technical personas, teams evaluating Modal migration.
Not recommended for: Teams seeking detailed architecture guidance (await blog post or see Stripe/Ramp).
Outlook: Abnormal AI's promised blog post should provide valuable architectural details. The 13% benchmark sets a new target for enterprise adoption.
Research by Ry Walker Research • methodology
Disclosure: Author is CEO of Tembo, which offers agent orchestration as an alternative to building in-house.