FurtherAI merges code without human approval using AI agents
An engineering team reports automating reviews, tests and merges, with a third of changes now deploying without direct human intervention.
Key points
- FurtherAI engineers merge five times more code daily while maintaining a revert rate below half a percent.
- AI agents now review ninety-nine percent of changes before humans, allowing automated merges for low-risk work.
- The organisation converted tribal knowledge into twenty-one automated Python checks that fail builds and print fixes.
FurtherAI states that up to a third of its code changes now merge automatically without human involvement. This capability relies on automated checks, AI reviews and passing tests.FurtherAI allows some code updates to deploy automatically because AI tools check them first.FurtherAI achieves autonomous merging for low-risk changes by integrating AI reviews with automated Python checks. The team reports this shift occurred in August, allowing low-risk modifications to deploy without a person reviewing the final merge.reported This represents a significant change in how engineering output is validated and released to production environments.
How we got here
The organisation began optimising its software shipping process in March by having agents write code while engineers steered them. By April, agents wrote tests alongside code, raising the inclusion of tests in changes from fifteen percent to ninety-one percent. In May, AI reviewed ninety-nine percent of changes before humans did, up from sixty percent. June saw parallelised tests run in the same seventeen minutes as before, handling twenty times the volume.
What is new
In July, the team used ClaudeTag to automate bug intake via Slack, filing sixty percent of tickets automatically. August introduced automated Python checks for tribal knowledge, converting twenty-one rules into build failures that agents can fix themselves. These checks enable low-risk changes to merge without human review. September added code channels where users describe changes to agents in Slack, opening pull requests collaboratively.
What to watch
The team notes that speeding up one step often piled work up at the next. They continue to optimise each step of the shipping process individually. The number of changes merged via code channels is not yet quantified. The organisation is still refining how multiple people steer the same agent in a single channel to add context and guide the output effectively.
Availability and limits
This workflow is internal to FurtherAI and not a public product. It requires specific tooling, including ClaudeTag and custom Python checks. The automated merge applies only to low-risk changes that pass all automated checks and AI reviews. High-risk or complex changes likely still require human judgment. The revert rate remains under half a percent, suggesting stability, but the exact scope of "low-risk" is not defined. Access to this specific pipeline is limited to their engineering staff.
Why it matters
- Developers can see a practical example of shifting human effort from routine reviews to high-judgment tasks.
- The data suggests AI agents can handle ninety-nine percent of initial code reviews without increasing revert rates.
- Converting tribal knowledge into automated checks allows agents to self-correct, reducing manual intervention.
- The five-fold increase in daily merges indicates potential for higher throughput when bottlenecks are removed.
Topics: AI agents, software engineering, code review, automation, FurtherAI
Updated
Then / now / next
Humans reviewed sixty percent of changes, and only fifteen percent included tests.
AI reviews ninety-nine percent of changes, and ninety-one percent include tests.
The team continues to optimise each step of the shipping process to prevent bottlenecks.
Story timeline
- Then Humans reviewed sixty percent of changes, and only fifteen percent included tests.
- 29 Oct 2026 Building an AI-native engineering org
- Now AI reviews ninety-nine percent of changes, and ninety-one percent include tests.
- Next The team continues to optimise each step of the shipping process to prevent bottlenecks.
The strongest doubt
Counterpoint from the sources
The author notes that speeding up one step often caused work to pile up at the next, requiring careful optimisation of the entire process.
For readers
Developers
- Agents now write code and tests, with humans steering the process.
- Low-risk changes can merge automatically if they pass AI reviews and checks.
Business
- Code merge volume increased fivefold while revert rates stayed under half a percent.
- Automated bug intake files sixty percent of tickets, reducing manual triage.
Researchers
- The study covers nine months of building an AI-native engineering organisation.
- It highlights the need to optimise the entire shipping process, not just coding.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
How many code changes does each engineer merge daily now?
The number of code changes each engineer merges daily has grown about five times since January.
What percentage of changes include tests now?
Ninety-one percent of changes now include tests, up from fifteen percent previously.
How are bug reports handled?
ClaudeTag automates bug intake from Slack, filing sixty percent of bug tickets and assigning them.
Connected stories
Reading receipt
- FurtherAI engineers merge five times more code daily while maintaining a revert rate below half a percent.
- AI agents now review ninety-nine percent of changes before humans, allowing automated merges for low-risk work.
- The organisation converted tribal knowledge into twenty-one automated Python checks that fail builds and print fixes.