Video by FINOS via YouTube
Erik Seliger (Lead Product Engineer at Sourcegraph) and Robert Lin (Lead Platform Engineer at Sourcegraph) join FINOS Zenith Project Lead Peter Smulovics to break down the mechanics of Agentic Batch Changes. They reveal how implementing a multi-layered agent architecture solves the high economic cost of pure "vibe coding" by combining deterministic scripting pipelines with localized, token-efficient AI coding models.
🗽 Catch Us in New York! Ready to scale your repository automation while defending corporate token budgets?
Join the enterprise DevOps community at OSFF New York on November 4–5, 2026.
🎟️ Register Now: https://hubs.ly/Q04n_bZL0
🔥 25% OFF DISCOUNT CODE: FINOSREDOSFF25
🕒 Timestamps:
0:00 Reminders: Call for Proposals and Exclusive OSFF New York Perks
1:35 Thank You to Our NY Sponsors: Red Hat, Moderne, and Temporal
1:59 Introducing Sourcegraph: Powering Code Intelligence for Global Banks
3:26 Session Kickoff: The Concept of Multi-Layered Agentic Code Modification
4:03 What is a Batch Change? Overcoming Tedious Multi-Repo Migrations
4:38 The Log4j Reality Check: Scaling Vulnerability Patches Across 25,000 Services
5:15 Why Developers Hate Manual Modernization Campaigns (and Cookie Bribes)
6:11 Scriptable Boundaries: The Limitations of Pre-AI Automation
6:38 Meet the Outer Loop: The Deterministic Project Manager Agent
7:15 Inner Loop Coordination: When to Program vs. When to Burn Tokens
8:48 The Cost Moat: Crafting Markdown Spec Files for Repeatable Scale
10:10 Before vs. After AI: Shifting From Manual Find-and-Replace to Self-Correction
11:09 Staged Validation: Running Contextual Code Search via DeepSearch
11:58 Scripting Efficiencies: Leveraging Automated Ecosystem Tools Natively
12:57 Approachable Migration: Watching Git Patches Evolve in Real-Time
14:09 The Prevention Moat: Stopping the Scope Creep of Automation Platforms
14:48 Bring Your Own Model: Swapping Outer Engines vs. Experience Tuning
16:15 Out-of-the-Box Setup: Integrating Native Sourcegraph MCP Connections
17:37 Deterministic Plus Agent Mix: The Quality Boost of Hybrid Recipes
19:10 Automating the Container Image Base via Outer Loop Orchestrators
19:59 Hook Blocks: Deploying Coding Agents to Automatically Fix Failed CI Pipelines
21:20 Deterministic Log Fetching: Pruning API Trajectories to Eradicate Waste
22:23 The Spectacular Scripting Flaw: How AI Wrote a Humongous Switch Case
23:26 Accessing the Beta: How Existing Enterprises Can Test Agentic Scale
24:32 The Ultimate Inflection Point: Flipping the Machine From a Money Burner to Magic
📊 The Problem: The Economic Nightmare of Unbounded "Vibe Coding" When global financial institutions attempt large-scale codebase modifications—such as upgrading framework dependencies across 25,000 active code services—relying on a single coding model introduces severe fragmentation. Traditional coding assistants immediately blow through millions of context-window tokens trying to execute code changes repo-by-repo. Furthermore, when these blind refactoring attempts trigger continuous integration (CI) failures, the model encounters a non-deterministic loop, spinning out out-of-bounds calculations, custom exceptions, and configuration drift that burns cash without hitting production.
🏗️ The Solution: The Multi-Layered Agentic Batch Architecture Sourcegraph delivers a structural framework that decouples project coordination from individual script executions:
The Outer Loop Project Manager: Running an orchestration agent governed by immutable system prompts to research, manage repository staging, and continuously audit diff outputs.
Deterministic Baseline Integration: Using the planning agent to write efficient, low-cost scripts for programmatic tasks (like straightforward find-and-replace or package manager updates) instead of burning tokens on every file.
Inner Loop Containment: Deploying localized coding engines (such as Cody or Cloud Code) strictly inside isolated base Docker images to handle ambiguous, complex code edge cases on demand.
⚙️ Why This Matters for Financial Engineering
Automated CI Hook Blocks: Injecting localized AI hooks directly into failing build logs, empowering inner loop agents to dynamically correct pipeline issues in real time based on compilation feedback.
Deep MCP Integration: Setting up customized Model Context Protocol (MCP) layers to bridge context lines, letting inner agents instantly ask the master outer orchestrator for repository-scale search data when hitting technical roadblocks.
🌐 More about FINOS: https://www.finos.org/
📧 Join our newsletter: https://www.finos.org/sign-up
🎙️ Listen to our Open Source in Finance Podcast: https://www.youtube.com/@FINOS/podcasts
LinkedIn: https://www.linkedin.com/company/finosfoundation
#FINOS #OSFFNewYork #Sourcegraph #BatchChanges #ZenithProject #DevOps #MultiAgent #TokenEfficiency #CIFixing #MCP