54% of engineering leaders say more than 60% of AI agent work needs major rework
An agent can produce a patch in minutes, but a human still has to understand and verify every change. In a Sourcegraph survey of 1,028 engineering leaders, 54% say more than 60% of AI work needs substantial manual rework before release. At the same time, 79% feel pressure to prove that AI pays off. That is why teams should measure agent adoption by review time and post-release failures, not just speed.
AI-processed from Sourcegraph Blog; edited by Hamidun News
Imagine this: an agent produces a patch in five minutes that updates a library across a dozen services. The reviewer still has to figure out where else this code is used, whether a service pinned to the old version will break, and which tests will catch the error. The Sourcegraph Big Code 2026 report measured this problem. 54% of respondents believe that more than 60% of AI-generated work requires substantial manual rework before release. In other words, more than half of leaders see an agent's output as a draft rather than a finished product. Writing code has become fast; understanding it still takes a long time.
The report is based on a survey by Dimensional Research. In September 2026, it polled 1,028 people responsible for development at companies with more than 250 developers. An important caveat: all figures are respondents' self-assessments, not measured metrics. In addition, Sourcegraph sells a code intelligence platform, so the report's conclusions point toward its products. Still, the data is interesting.
Pressure to prove ROI and rework
79% of respondents at companies that use AI feel moderate, significant, or strong pressure to show a return on their AI investment. At the same time, 58% call scaling AI and agents a priority for the next twelve months. That means there will be more agents, and therefore more other people's changes to review.
The authors also cite the DORA research: organizations that use AI more tend to ship changes faster and rate their product more highly, but their delivery is less stable. The practical advice follows: don't measure the success of adoption by speed alone. Track how much effort goes into reviewing each change and how often problems appear after release.
Understanding someone else's code is half the job
An agent can make a change quickly, but the reviewer has to trace dependencies across repositories, check local rules, and work out which tests will catch a regression. This can take hours of a senior engineer's time, plus that of colleagues who know the affected systems.
According to the survey, 65% estimate that reading, searching, and understanding existing or proposed code takes more than a quarter of team time, and 25% say it takes more than half. To put that in scale: if an organization of 1,000 developers spends a quarter of its time on this, that equals the work of 250 people. 51% of respondents have at least 1,000 repositories, and the question "what depends on this code?" easily extends beyond the repository where the change started. 41% say engineers regularly rewrite things that already exist because they can't find them.
The sharpest example is a critical vulnerability in a dependency. Asked how long it would take to find and fix all affected locations, 54% answered "days or longer" or "we couldn't do it." These are hypothetical estimates, not measurements of actual fixes. Among the main difficulties with large-scale changes, respondents most often named understanding dependencies and impact (43%), keeping changes consistent across repositories (42%), and testing (42%).
Agents lack shared context
Only 10% say all of their agents see the full picture of code, dependencies, and history. For the rest, context is shared only partially or is fed to each agent separately. Another 33% of companies using AI admit that developers must explain the architecture, dependencies, and agreements in detail before an agent can start on a new task.
Sourcegraph offers a solution: a server through which compatible agents such as Claude Code, Codex, and Cursor search the organization's indexed code during a task. It respects existing repository access permissions, and administrators control who can use which tools. The engineer sees the retrieved code next to the proposed change, and when something fails, it's clear what went wrong: the agent didn't find the right code, misread an exception, or made an implementation error. For mass upgrades and migrations, the company also offers a separate tool for coordinating changes across repositories.
What follows from this
The last figure stands out: 77% expect that AI-written changes will be merged with less human oversight in the future. To decide which changes can be given that kind of autonomy, you need to see where agents succeed and where engineers have to step in.
For teams, the takeaway is simple. First, measure review time and the share of changes that cause problems after release, not just release speed. Second, invest in code search and shared context for agents, or the savings from writing code will be lost in review. Third, treat vendor surveys critically: respondents were estimating by feel, and the sample covers only large companies.
- Nov 21, 2026VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
Key Questions Answered
What are the key details on Pressure to prove ROI and rework?
79% of respondents at companies that use AI feel moderate, significant, or strong pressure to show a return on their AI investment. At the same time, 58% call scaling AI and agents a priority for the next twelve months. That means there will be more agents, and therefore more other people's changes to review. The authors also cite the DORA research: organizations that use AI more tend to ship changes faster and rate their product more highly, but their delivery is less stable.
What are the key details on Understanding someone else's code is half the job?
An agent can make a change quickly, but the reviewer has to trace dependencies across repositories, check local rules, and work out which tests will catch a regression. This can take hours of a senior engineer's time, plus that of colleagues who know the affected systems. According to the survey, 65% estimate that reading, searching, and understanding existing or proposed code takes more than a quarter of team time, and 25% say it takes more than half.
What are the key details on Agents lack shared context?
Only 10% say all of their agents see the full picture of code, dependencies, and history. For the rest, context is shared only partially or is fed to each agent separately. Another 33% of companies using AI admit that developers must explain the architecture, dependencies, and agreements in detail before an agent can start on a new task. Sourcegraph offers a solution: a server through which compatible agents such as Claude Code, Codex, and Cursor search the organization's indexed code during a task.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.