Loading Runway...
Loading Runway...
Evidence-backed analysis across 20 specific tasks. Capability claims sourced from peer-reviewed research, independent benchmarks, and industry data. Adoption rates tracked by industry and company size.
At a glance
Early Signal intelligenceTasks tracked
Signals in database
Intelligence confidence
Last updated
AI Exposure
Defensibility
Avg Capability
20/20 tasks with evidence
Avg Deployment
302 evidence sources
What's changing for Software Engineers
Hiring volumes for software engineers contracted sharply in 2023–2024 following mass layoffs across big tech, but stabilised in late 2024. Demand is now bifurcating: engineers who can work effectively with AI coding tools (Copilot, Cursor, Claude) and ship at higher throughput are commanding salary premiums, while pure implementation roles at junior and mid levels face real compression. Backend and infrastructure engineers with distributed systems depth remain in short supply. Full-stack generalists without clear systems or product instincts are experiencing the most pricing pressure. Companies rebuilding headcount post-freeze are hiring senior engineers at a 2:1 ratio over junior hires — a structural shift that is likely to persist. Security-aware engineering is becoming a baseline requirement, not a specialism, following a run of high-profile supply-chain incidents. AI is absorbing first-draft code generation and boilerplate; the premium is shifting toward system design judgment, debugging complex distributed failures, and owning production reliability. Engineers who avoid AI tooling are falling behind on raw throughput benchmarks in hiring assessments.
Synthesised by claude-sonnet-4-6 · refreshed May 21, 2026
Capability dimensions
How the dimensions of this role are being reshaped by AI · top 8 by weight
Implementation Quality
System & Architecture Design
Technical Fluency
Problem Decomposition
Quality Assurance & Review
Technical Documentation
Estimation & Scoping
Reliability & Operational Excellence
Market Context
41% of all code written in 2025 is AI-generated (Stack Overflow, 49K developers). Traditional programmer employment declined 27.5% between 2023–2025 per US BLS data. Mid-level coding tasks (boilerplate, unit tests, bug fixes) are now largely AI-assisted. Demand is bifurcating: AI-fluent engineers commanding premiums; traditional stack developers seeing reduced demand.
Source: Based on Stack Overflow Developer Survey 2025, BLS Occupational Outlook 2025, GitHub Copilot adoption data, and METR July 2025 RCT findings on AI-assisted development.
Task breakdown
Top 3 per pressure tier · expand for the full list
Medium automation pressure · 6
Unit & Integration Testing
Event-based autonomous agents demonstrate expanded computer use capabilities beyond reactive prompting, requiring cross-tool integration and signal detection.
Technical Documentation
While the tool automates documentation restructuring, it enhances rather than replaces technical writing roles and lacks enterprise adoption metrics.
CI/CD Pipeline Management
Gartner Leadership recognition with enterprise-scale deployment indicates incremental improvement in real-world code generation reliability and integration.
Dependency Management
AI-powered dependency scanning tools (Dependabot, Snyk, Renovate) can automatically identify outdated dependencies, flag known vulnerabilities from CVE databases, and generate pull requests for version updates. GitHub re…
Security Review & Hardening
Mythos can discover security vulnerabilities faster and more comprehensively than entire teams of human security professionals
Low automation pressure · 14
Feature Implementation
Census BTOS (period 100): 42.9% of U.S. establishments in NAICS 51 (Information) reported using AI in the last two weeks (5/4/2026 - 5/17/2026). Sector-level measure — not mapped to a single Runway task. Sample frame ~1,…
Release & Deployment Management
Large-scale production deployment of AI for code generation across 70% of codebase demonstrates sustained capability for real-world software development at enterprise scale.
Code Review
Role Defensibility Profile
Higher = harder to automate
Task-Level Analysis — 20 Tasks
Write production code to implement new features and functionality — translating requirements and designs into working, tested, and deployable software.
Highest Exposure Areas
Hands-On Technical Execution
41% of code written in 2025 is AI-generated. The defensible technical work is system architecture, novel problem-solving, and integration of AI tools — not execution of known patterns. Standard technical execution is being absorbed at an accelerating rate.
Meetings / Coordination / Scheduling
Calendar AI and agentic scheduling tools already handle meeting coordination. The coordination value that remains human is the nuanced political navigation — and that erodes as AI gains organisational context.
Writing / Summarising / Documentation
GPT-5 Deep Research and Claude already produce publication-quality reports, emails, and documentation. By 2027, AI writing assistants will handle first-draft creation for virtually all standard business documents with minimal human input.
Strongest Defenses
Hands-On Technical Execution
41% of code written in 2025 is AI-generated. The defensible technical work is system architecture, novel problem-solving, and integration of AI tools — not execution of known patterns. Standard technical execution is being absorbed at an accelerating rate.
Decision-Making Under Uncertainty
This remains one of the most defensible task categories — AI struggles with genuine novelty and accountability. The erosion condition: as AI decision-support tools become standard, the bar for what counts as 'genuine uncertainty' rises, and roles that mostly execute defined playbooks lose this protection.
Domain Specialist Judgement
Deep domain expertise is the most durable protection — but it degrades when AI is trained on sufficient domain-specific data to match pattern recognition. The erosion condition: the more codifiable your expertise, the faster this protection erodes. Truly novel, context-dependent judgement remains human-critical.
See software engineers by industry
Same role, different industry-specific exposure profiles.
Pick another role to see a side-by-side AI disruption comparison. The URL you land on is shareable.
Live signals
Real-time AI signals affecting this role
Compare roles
See how other roles compare
What this means for software engineers
The role-average exposure profile above is built on early signals — directionally useful but not yet corroborated across independent sources. Your specific task mix and tooling matter more than the role average here. Get a personal task-level breakdown rather than relying on the headline number.
How we build role intelligence
Runway maintains an atomic task taxonomy (20 tasks tracked for Software Engineer) anchored to O*NET occupational data. Per-task signals enter through tier-graded connectors (peer-reviewed papers, statutory labour data, vendor benchmarks, preprints) and pass through the Sentinel auditor — every claim is rubric-scored, cross-checked, and confidence-graded before it can affect a role page. The narrative and task breakdown above are computed from that ledger; nothing is synthesised from first principles. See /methodology for the full pipeline.
Confidence level: Early Signal — based on 1 validated signal for this role across the Sentinel-graded sources we track.
Monitoring & Observability
Autonomous optical network system provides monitoring and observability across distributed AI training lifecycle with ~98% task completion rate
Cognition's $26B valuation and large funding round validate Devin's autonomous code generation capabilities in production, confirming sustained high performance in code automation.
Debugging & Troubleshooting
On SWE-bench Verified, which consists of real bug reports from popular open-source Python repositories, frontier LLMs resolve 40-50% of issues autonomously. These are genuine debugging tasks requiring diagnosis from issu…
Performance Optimisation
DeepSWE benchmark shows GPT-5.5 achieving significantly higher autonomous code completion accuracy than previous models, indicating substantial improvement in real-world coding task performance.
API Design
New realtime voice models with multi-dimensional capabilities (transcription, translation, reasoning) demonstrate measurable advancement in autonomous speech-based task handling deployed at scale through OpenAI's API.
Technical Debt Management
AI-assisted code generation that automates code writing for developers and technical teams
Technical Estimation
Agentic AI workflows powered by Codex can automate development tasks for enterprise technical teams at scale
Requirements Clarification
AI systems can filter content based on regional requirements
Database Schema Design
LLMs can generate normalised relational schemas, write migration scripts, and suggest indexing strategies for standard data modelling patterns. Eloundou et al. classify database design tasks as having moderate-to-high LL…
Incident Response & On-Call
Mean time to recovery reduced by 50% through AI-assisted incident handling
Architecture & System Design
AI system can independently complete software engineering tasks
Cross-Team Technical Collaboration
Agentic AI workflows powered by Codex can automate development tasks for enterprise technical teams at scale
Onboarding & Knowledge Transfer
Can automate research workflows for analysts and knowledge workers
Capability Evidence
DARPA's demonstrated AI systems successfully identified injected vulnerabilities across 54 million lines of actual production code, showing significant advancement in automated code analysis and bug d...
AI agents can handle routine software implementation tasks at a level threatening mid-level software engineering roles
AI agents can autonomously debug code in software codebases
Deployment by Industry
Diagnose and fix bugs in existing code by reproducing issues, analysing logs and stack traces, isolating root causes, and implementing targeted fixes.
Capability Evidence
LLMs can generate reliable EDA (Electronic Design Automation) code for chip design tasks without requiring iterative tool-in-the-loop debugging
Can detect and repair security vulnerabilities in code, which is a specific type of debugging focused on security issues
AI agents with hardware-in-the-loop integration capabilities can automate debugging of embedded systems
Deployment by Industry
Review pull requests from teammates for correctness, readability, performance, security, and adherence to codebase conventions — providing constructive feedback and approving changes.
Capability Evidence
Cognition's $26B valuation and large funding round validate Devin's autonomous code generation capabilities in production, confirming sustained high performance in code automation.
Demonstrated production use case showing Codex accelerating substantive code review tasks, indicating improved understanding of code context and quality assessment.
Fine-tuned models now deliver contextual code review feedback, extending code analysis capabilities into pedagogical domains previously requiring human instructors.
Deployment by Industry
Write and maintain automated tests — unit tests, integration tests, and end-to-end tests — to verify correctness and prevent regressions.
Capability Evidence
Event-based autonomous agents demonstrate expanded computer use capabilities beyond reactive prompting, requiring cross-tool integration and signal detection.
LLMs demonstrate strong capability in generating unit tests from existing code. The Peng et al. GitHub Copilot study found that AI-assisted developers produced code with equivalent test pass rates to ...
Mistral Small 4's coding capabilities can assist with writing unit and integration tests
Deployment by Industry
Design system architecture for new services, features, or infrastructure — making decisions about data models, service boundaries, API contracts, and scalability patterns.
Capability Evidence
AI system can autonomously apply security patches to enterprise codebases
AI control architecture can perform deterministic, safety-constrained automation of software development processes beyond code generation alone
Multi-agent AI system can automatically convert process sketches into executable simulation models
Deployment by Industry
Design REST, GraphQL, or RPC API interfaces — defining endpoints, request/response schemas, error handling patterns, versioning strategies, and authentication flows.
Capability Evidence
New realtime voice models with multi-dimensional capabilities (transcription, translation, reasoning) demonstrate measurable advancement in autonomous speech-based task handling deployed at scale thro...
AI agent system can run entirely on local hardware without external API dependencies
AI system can perform chat conversations in an OpenAI-compatible API format
Deployment by Industry
Write and maintain technical documentation — architecture decision records, API docs, runbooks, README files, and inline code documentation.
Capability Evidence
While the tool automates documentation restructuring, it enhances rather than replaces technical writing roles and lacks enterprise adoption metrics.
AI systems can automatically generate software documentation from source code using prompt-driven techniques
Technical documentation and writing tasks are among the highest-exposure categories in the Anthropic Economic Index analysis. Writing-related tasks represent the single largest category of professiona...
Deployment by Industry
Profile, benchmark, and optimise application performance — identifying bottlenecks in code, queries, network calls, and rendering paths, then implementing targeted improvements.
Capability Evidence
DeepSWE benchmark shows GPT-5.5 achieving significantly higher autonomous code completion accuracy than previous models, indicating substantial improvement in real-world coding task performance.
LLMs can identify common performance anti-patterns in code (N+1 queries, unnecessary re-renders, inefficient algorithms) and suggest standard optimisations. However, the SWE-bench evaluation framework...
LLMs can generate optimized Triton operators for GPU kernels, though still requiring trial and error according to the research
Deployment by Industry
Review code and infrastructure for security vulnerabilities — injection attacks, authentication flaws, data exposure risks — and implement mitigations and security best practices.
Capability Evidence
Autonomous contextual analysis for complex fraud/impersonation detection in production email security represents capability advancement in reasoning over nuanced security signals.
Codex Security AI agent demonstrates autonomous vulnerability detection and threat model generation, but requires human validation of attack paths, indicating partial automation of security analysis r...
Specialized cybersecurity variant deployed to verified professionals indicates incremental improvement in domain-specific reasoning for security analysis tasks.
Deployment by Industry
Work with product managers and designers to clarify ambiguous requirements, identify edge cases, surface technical constraints, and refine specifications before implementation begins.
Capability Evidence
The Anthropic Economic Index found that tasks requiring stakeholder negotiation, ambiguity resolution through dialogue, and understanding unstated assumptions are among the least automated professiona...
The IMF finds that approximately 40% of global employment is exposed to AI, with up to 60% in advanced economies. For knowledge work tasks like Requirements Clarification, the study estimates 21% of t...
Eloundou et al. classify requirements gathering and stakeholder communication tasks as having low LLM exposure. While models can help structure requirements documents and generate user story templates...
Deployment by Industry
Collaborate with engineers on other teams to align on shared interfaces, resolve integration issues, coordinate migrations, and ensure consistent technical standards.
Capability Evidence
The Anthropic Economic Index identifies cross-functional coordination, negotiation, and relationship management as the professional task categories with the lowest LLM automation potential. Cross-team...
Agentic AI workflows powered by Codex can automate development tasks for enterprise technical teams at scale
AI coding agents can automate part of advanced technical job functions in AI development roles
Deployment by Industry
Onboard new team members to the codebase, conduct pairing sessions, explain system context and historical decisions, and ensure institutional knowledge is shared effectively.
Capability Evidence
AI copilot can embed industry-specific knowledge to assist with domain-specific enterprise tasks
AI system can traverse knowledge graphs to understand and navigate complex codebase relationships for maintenance decisions
AI systems can automate industrial operations tasks that currently require specialized knowledge of industrial protocols
Deployment by Industry
Configure, maintain, and improve continuous integration and deployment pipelines — build scripts, test automation, deployment workflows, and environment management.
Capability Evidence
Gartner Leadership recognition with enterprise-scale deployment indicates incremental improvement in real-world code generation reliability and integration.
Demonstrates practical tool integration for domain-specific data access but no deployment scale metrics provided.
Managed Agents API demonstrates improved autonomous agent orchestration and deployment automation, reducing engineering overhead for multi-step agent workflows.
Deployment by Industry
Respond to production incidents — triaging alerts, diagnosing issues under time pressure, implementing hotfixes, coordinating with affected teams, and writing post-mortems.
Capability Evidence
Anthropic's study of real-world Claude usage across millions of professional conversations found that tasks related to Incident Response & On-Call represent a significant category of AI-augmented work...
The Anthropic Economic Index found that real-time decision-making tasks requiring organisational context, live system state awareness, and cross-team coordination are among the least automated by curr...
AI can augment incident response workflows for security teams
Deployment by Industry
Design and evolve database schemas — table structures, indexes, constraints, migrations, and query optimisation — balancing normalisation, performance, and maintainability.
Capability Evidence
Anthropic's study of real-world Claude usage across millions of professional conversations found that tasks related to Database Schema Design represent a significant category of AI-augmented work. The...
The Anthropic Economic Index found that data-related tasks including schema design and query optimization are common professional uses of LLMs. Models perform well on standard schema patterns and can ...
Cognizant and Oxford Economics analysed 18,000+ tasks across industries and found that Gen AI will impact 90% of jobs but fully displace very few. For tasks like Database Schema Design, the study esti...
Deployment by Industry
Estimate development effort, complexity, and risk for proposed features and projects — breaking down work into tasks and providing time and resource estimates to product and leadership.
Capability Evidence
The IMF finds that approximately 40% of global employment is exposed to AI, with up to 60% in advanced economies. For knowledge work tasks like Technical Estimation, the study estimates 24% of task co...
Agentic AI workflows powered by Codex can automate development tasks for enterprise technical teams at scale
AI can perform multi-step development tasks autonomously
Deployment by Industry
Evaluate, update, and manage third-party dependencies — assessing security advisories, compatibility, licensing, and upgrade paths for libraries and frameworks.
Capability Evidence
AI system can autonomously manage dependency updates across enterprise repositories
Automate building-grid co-simulation for energy management
MIT Sloan Management Review's annual survey of 3,000+ managers found that only 10% of organizations report significant financial value from AI deployment, despite widespread experimentation. For tasks...
Deployment by Industry
Identify, document, and systematically address technical debt — refactoring legacy code, improving test coverage, and modernising outdated patterns without breaking existing functionality.
Capability Evidence
LLMs can perform bug detection workflow automation by analyzing code segments without test coverage
AI systems can perform legacy code modernization and memory safety refactoring, a task previously requiring specialized human engineers
LLMs can identify code smells, duplicated logic, overly complex functions, and common anti-patterns through static analysis augmentation. The SWE-bench evaluations demonstrate that models can successf...
Deployment by Industry
Instrument applications with logging, metrics, and tracing — configuring dashboards, alerts, and health checks to ensure production systems are observable and issues are detected early.
Capability Evidence
Identification of confident-but-wrong autonomous behavior in production observability agents suggests current computer_use capability has significant safety gaps, requiring downward adjustment from th...
LLMs can generate monitoring configurations, write alerting rules, create dashboard definitions, and instrument code with tracing and metrics collection. The Anthropic Economic Index identified infras...
The IMF finds that approximately 40% of global employment is exposed to AI, with up to 60% in advanced economies. For knowledge work tasks like Monitoring & Observability, the study estimates 27% of t...
Deployment by Industry
Coordinate and execute software releases — managing feature flags, rollout strategies, canary deployments, rollback procedures, and post-deployment verification.
Capability Evidence
Radiology workflow agent deployment demonstrates multi-step autonomous task orchestration in a specialized healthcare context, validating existing computer_use capability without clear evidence of imp...
Large-scale production deployment of AI for code generation across 70% of codebase demonstrates sustained capability for real-world software development at enterprise scale.
Gemini deployment to 4M vehicles represents significant real-world deployment scale but does not indicate a breakthrough in underlying AI capability, only market penetration.
Deployment by Industry