Engineering platforms for understanding repetitive work and engineering investment
How Weave, Jellyfish, Swarmia, LinearB and DX show where engineering effort goes across features, bugs and maintenance without relying on timesheets.
On this page
Weave shows where engineering effort goes by classifying every merged pull request as Feature, Bug or KTLO (keeping the lights on). It then weights each PR by the expert effort its code represents. Weave's developer surveys add what code cannot show, including how engineers rate their focus time, context switching and tooling friction.
The approaches compared here use shipped work, issue-tracker structure, effort estimates and survey answers as proxies for work allocation. Code-derived categories do not by themselves measure the hours people spent on repetitive or creative work; direct time studies and developer feedback can add that context. This guide is published by Weave. It compares how Weave, Jellyfish, Swarmia, LinearB and DX build those proxies, based on each vendor's current website as reviewed on November 14, 2026.
Count effort in each category, not PRs
Take an illustrative sprint in which a team merges 40 dependency updates and one new feature. By PR count, the sprint looks like nearly all maintenance. Suppose the feature required schema changes, new service logic and updates across several files. An effort-weighted view may then show that it absorbed a large share of the sprint's engineering work. Allocation decisions depend on that weighted share, such as whether a team is short of roadmap time.
Weave's code-output model supplies the weight. It estimates the effort an expert engineer would need to complete each change, combining an LLM, static analysis and a custom model trained on engineer-labelled PRs. Each score comes with a plain-language explanation. A score of 3.0 means three hours of modeled expert effort. It does not record when anyone worked or for how long.
How Weave categorizes engineering work
Feature, Bug and KTLO
Weave's AI categorization assigns each PR one of three types:
- Feature: new functionality.
- Bug: repairs to broken behavior.
- KTLO: other maintenance, including refactoring, performance improvements, documentation, tests, moving code and deployment work.
Users can override any label. Where the model misreads a change, the allocation then reflects the team's own judgment.
Group PR output by component
Weave also labels PRs by component, starting with frontend, backend and infrastructure. Teams can add their own components, such as a payments service or a data pipeline. When a PR touches several components, component charts split its output evenly across them. Custom components apply only to PRs merged after the component is created, so set them up before the period you want to analyze.
Find recurring maintenance within KTLO
KTLO is not a synonym for repetitive work. A database migration or a performance fix under production load can be among the hardest changes a team ships. Repetitive work tends to appear as a pattern:
- Many low-effort KTLO PRs in the same component.
- High code turnover in those files.
- Repeated review rounds on similar changes.
Weave's custom queries let a manager investigate that pattern in plain English or SQL. Queries can draw on PR categories, output, AI contribution, turnover and review data. The result can be saved as a dashboard widget that refreshes about hourly. The same query shows whether coding agents already carry much of the routine maintenance in a given component.
Pair the work data with what engineers report
Code categories show what shipped, but they cannot show which work felt draining. Weave's developer experience surveys are available to every organization. The default bank has 12 questions in four sections:
- Delivery and tooling friction.
- Cognitive load and documentation.
- Product specification and focus time.
- Culture and collaboration.
Admins can add custom questions, such as asking which recurring tasks engineers would most like automated. Responses are anonymous unless an admin enables identifiable mode, and individual answers are then visible only to admins. Results appear as distributions and trends across survey cycles, beside the engineering data. A team that reports long waits for review can therefore be checked against its measured review turnaround in the same period.
How work-allocation approaches differ
Scroll horizontally if needed →
| Platform | How work is grouped | Strongest use |
|---|---|---|
| Weave | PR-level Feature, Bug and KTLO types plus customizable components, weighted by estimated expert effort, with overrides | A code-grounded view of new work versus maintenance, including how much AI contributes to each |
| Jellyfish | Investment reporting connected to AI usage, cost and capacity | Reporting engineering investment to executives and finance across initiatives |
| Swarmia | Investment balance, initiatives and capitalization alongside developer surveys | Team-level balance discussions within a continuous improvement program |
| LinearB | Delivery metrics with software-factory context and gitStream PR automation | Automating repetitive review and merge steps once they are identified |
| DX | Developer experience research with complexity-weighted throughput | Survey-led diagnosis of friction and focus time |
Jellyfish and Swarmia are strong choices when allocation reporting is organized around initiatives and finance. LinearB helps automate PR steps a team has already found to be repetitive. DX brings depth in developer research. Weave adds an allocation grounded in the code of each merged PR, weighted by effort and correctable PR by PR. Leaders can inspect the maintenance PRs behind an allocation change before deciding whether to adjust roadmap commitments.
Choose automation and staffing changes from the allocation
An effort-weighted split leads to specific decisions:
- Move a recurring class of maintenance to coding agents.
- Invest in tooling where survey friction and code turnover point to the same component.
- Protect focus time for a team whose feature share has fallen.
OpenZeppelin estimated that detecting issues early with Weave probably saved four weeks of engineering time. Intellistack connected GitHub, Cursor and Claude Code to build historical views of AI adoption for board reporting. Wooly, Weave's research agent, answers follow-up questions about a category or component with links to the underlying PRs.
Classify last quarter's merged work
Connect your repositories and review last quarter's Feature, Bug and KTLO split by estimated effort. Then sit with each team lead to correct misclassified PRs and flag the maintenance that felt repetitive. Run one survey cycle in parallel so the team's account sits beside the code record. Book a Weave demo to see the categorization on your own repositories.
See the work behind your AI spend
Compare code output, AI cost and quality, then inspect the work behind the trend.
Related guides
Engineering analytics tools with customizable dashboards
Build views of review time, delivery and developer feedback.
The Intelligence Layer for Software Factories
How code, agent sessions and AI spend connect to engineering decisions.
Best engineering analytics tools
Compare productivity, review quality, delivery metrics and AI cost reporting.