Jobs / Software engineering
Can agents do software engineering work?
Build, test, and maintain software, from issue fixes to whole applications.
- US jobs
- 2,150,380 (1.4% of employment)
- Typical wage
- $129K a year (employment-weighted mean of occupation medians)
- Occupations
- 7 in O*NET 31.0
- Evidence
- 9 direct benchmarks, 8 also relevant, 9 context only
The answer today
Most tasksNo change since 10 Oct 2025.
Rule: Agents complete 60% or more of the tasks (or projects) on two direct benchmarks from independent institutions.
The evidence that sets this step
- SWE-Bench Pro V2 by Scale AI and Reflection: 99.4% of tasks completed · Opus 5 (Claude Code) xhigh, 22 Sep 2026
- Aider Polyglot leaderboard by Aider: 88% of tasks completed · gpt-5 (high), 23 Aug 2025
Next step: Nearly all tasks needs 90% or more on two direct benchmarks from independent institutions.
How much of the work is measured
47% of the work in this family is measured by direct benchmarks. Some of the work is measured (20% to 50%). On 10 Oct 2025: 29%.
Work = the O*NET 31.0 work activities of the family's occupations, weighted by how important each task is and by US employment. A benchmark measures an activity when its tasks make the agent do it and its grading checks it. Method.
| Work activity (O*NET) | Share of the work | Direct benchmark |
|---|---|---|
| Evaluate designs, specifications, or other technical data | 15.5% | Measured |
| Communicate with others about specifications or project details | 11.8% | Measured |
| Design computer or information systems or applications | 9.8% | Measured |
| Direct scientific or technical activities | 6.8% | Not measured |
| Advise others on the design or use of technologies | 6.3% | Not measured |
| Develop technical specifications for products or operations | 6.2% | Not measured |
| Coordinate with others to resolve problems | 5.7% | Not measured |
| Supervise personnel activities | 5.4% | Not measured |
| Develop operational or technical procedures or standards | 4.7% | Not measured |
| Monitor operation of computer or information technologies | 3.9% | Not measured |
Our reading
Agents complete most software engineering tasks on direct benchmarks. On SWE-Bench Pro V2, the best entry resolves 99.4% of 642 repository issues. On Aider Polyglot, the best model solves 88% of 225 coding exercises. Fresh issues show a lower rate: 64.5% on SWE-rebench, where every model runs in one common scaffold. Limits: SWE-Bench Pro V2 was near its ceiling at launch, and Aider Polyglot uses small exercises and has no entries since late 2025. No direct benchmark measures advising others on the use of technologies or directing technical activities. Whole projects are not measured: CodeClash gives an Elo score, not a completion rate.
Evidence: SWE-Bench Pro V2, Aider Polyglot leaderboard, SWE-rebench, CodeClash by Aider, Nebius, Princeton University, Reflection, Scale AI, and Stanford University. Written 24 Sep 2026.
Benchmarks for this work
Direct evidence
Work benchmarks mapped primary to this family. They set the step when they report how many tasks agents complete.
- 1385 EloClaude Sonnet 4.5 · 3 Nov 2025Unrated
- 64.8%Claude Code + Opus 5.5 (max) · 6 Oct 2026Strong
Also relevant
These benchmarks cover part of this family's work, or have no result yet. They do not set its step.
- BIRD-Interactby HKU and Google Cloud22.7%MERIT + GPT-5.4 (GPT-4o user simulator) · 22 May 2026Emerging
- 96.7%Genloop's Sentinel Agent v2 Pro · 1 Mar 2026Saturated
- 84.7%NexAU-AHE + GPT-5.5 · 23 Apr 2026Retired
Context only
These benchmarks span many occupations, or they test how agents behave. They give context but do not set the step.
- 42.9%TTE-MatrixAgent + DeepSeek-V3.2 · 10 Nov 2025Emerging
Retired
Their maintainers or a major evaluator stopped using them as a frontier measure. They do not count.
- 79.2%Sonar Foundation Agent + Claude 4.5 Opus · 5 Dec 2025Retired
Functions: Feature developmentShipping pull requestsAsking for clarificationIssue resolutionData engineering and SQLCode reviewCodebase evolutionLong-horizon autonomyVulnerability researchVulnerability patchingProfessional deliverablesHardware design and debuggingTest integrityMCP tool useWeb researchEnterprise workflowsFreelance projectsTerminal operationsML engineering
What we do not know
- Benchmarks measure tasks, not whole jobs. A job is a bundle of tasks, judgment, and relationships.
- Capability is not adoption. A task that agents can do in a benchmark may still be done by people at work.
- Most results come from single runs. Reliability over many runs is often unknown.
Occupations in this family
Sorted by US employment (BLS OEWS May 2025). O*NET sub-occupations share the employment figure of their SOC group. Work measured = the share of the occupation's O*NET work that direct benchmarks test.
| Occupation | SOC | US jobs | Work measured |
|---|---|---|---|
| Software Developers | 15-1252.00 | 1.7M | 49% |
| Blockchain Engineers | 15-1299.07 | (435K group) | 22% |
| Software Quality Assurance Analysts and Testers | 15-1253.00 | 187K | 34% |
| Video Game Designers | 15-1255.01 | (113K group) | 44% |
| Web and Digital Interface Designers | 15-1255.00 | 113K | 44% |
| Computer Programmers | 15-1251.00 | 92K | 52% |
| Web Developers | 15-1254.00 | 70K | 32% |
How to cite
Credit the original work first: the benchmarks listed on this page, and their institutions.
Then, if you used this page:
Can Agents Work. "Can agents do software engineering work?." https://canagentswork.com/jobs/software-engineering/ (accessed 2026-10-11). CC BY 4.0.@misc{caw-jobs-software-engineering,
title = {{Can agents do software engineering work?}},
author = {{Can Agents Work}},
year = {2026},
howpublished = {\url{https://canagentswork.com/jobs/software-engineering/}},
note = {Accessed 2026-10-11. CC BY 4.0}
}