Cutting Costs and Preventing Hallucinations with Multi-Agents Learned from Uber's uReview System
TuBrief 편집팀
2026년 9월 8일
0
Computing/Software원본 영상을 바탕으로 AI의 도움을 받아 작성했습니다. 원본 영상이 기준입니다.
커뮤니티의 다른 글
댓글 (0)
Log in to leave a comment
아직 작성된 글이 없습니다
원본 영상을 바탕으로 AI의 도움을 받아 작성했습니다. 원본 영상이 기준입니다.
Log in to leave a comment
아직 작성된 글이 없습니다
Feeding raw git diffs and entire source code directly into a legacy monolith triggers a massive cost explosion. The uReview platform unveiled by Uber in 2025 automatically analyzed 90 percent of the 65,000 changes generated weekly across six monorepos. However, if you feed models into an unrefined codebase with unclear module boundaries without cleaning, token costs spike up to 6 times. Models that fail to understand global factory classes demand irrelevant null pointer checks and induce hallucinations. The moment useless comments pile up, developers turn off notifications and fall into alert fatigue.
You need to build a preprocessing engine that uses Python's standard ast module to extract only the changed function signatures and affected local variables. After refining the static metadata, hash comparisons wipe out functional modifications unrelated to the changes. Then, it reverse-parses only the signature and actual modified statements of the lowest-level function node containing the changed lines to create a lightweight JSON payload.
Running a benchmark with 100 legacy modules shows a stark difference. Raw file injection consumes 42,000 tokens and $0.273 per PR. On the other hand, using AST preprocessing JSON compression drops it down to 6,100 tokens and $0.068 per PR. API costs decrease by 47.3 percent, and the hallucinatory false-positive rate drops to 9.4 percent.
Bundling security agents and business logic agents into an interactive group chat causes them to bounce outputs back and forth, falling into an endless ping-pong loop. Analyzing a single PR takes tens of minutes, and API costs skyrocket. You must not let agents communicate directly with each other. You need to build a finite state machine framework leveraging Pydantic schemas and LangGraph.
Each specialized agent receives only the central pipeline state and returns results judged according to its domain rules strictly in a predefined model format. Define data classes using Pydantic that enforce identifiers, file paths, severity levels, and confidence scores for agent outputs. Implement conditional edges that forcibly cut off the state machine when the iteration count reaches 2 or when the confidence of all comments converges to 0.85 or higher.
According to the OpenTelemetry GenAI Semantic Conventions standard, each agent's execution must be wrapped in a unique span, and token usage should be left in a distributed tracing backend. Applying this structure reduces the debugging time tech leads used to waste catching agent malfunctions by more than 6 hours per week.
According to operational data from Google's internal static analysis tool Tricorder, no matter how accurate a tool's critique is, if developers feel it is not worth fixing, hostility only grows. Checkers with an effective false positive rate exceeding 5 percent are immediately expelled. To make over 67 percent of automated review comments accepted, you need a gradual rollout starting from low-risk modules and expanding step-by-step.
Run shadow mode for 3 weeks, hiding comments in layers with no side effects, such as DTO validation layers and pure functions. After confirming 85 percent precision, unleash inline comments on backend code across three domain squads and track the acceptance rate. Collect feedback logs on a weekly basis to run a feedback loop script that automatically drops high-false-positive rules with acceptance rates below 67 percent from the whitelist and throws them into an isolation list. Embedding this whitelist verification script into the pipeline quickly filters out useless rules that annoy developers, allowing you to hit a 70 percent review system acceptance rate within the team.