Agent Performance in Large Codebases: Context Matters

Great piece (refer below) — your framework for context efficiency maps exactly to a problem I've been solving hands-on for a large .NET case management system (25 repos, 16 databases, FACTS framework). Where our approaches diverge: Sourcegraph MCP gives agents better search over raw code. I built a system that gives agents pre-distilled architectural knowledge — and I think the token economics actually favor distillation. Here's what I mean: an Ollama-powered seeder walks every .cs/.cshtml/.dsm file across all 25 repos, runs role-aware prompts (Controller → extract ENDPOINTS, AUTH, INJECTED_DEPS; Startup → extract MIDDLEWARE, DI_REGISTRATIONS, etc.), and stores structured summaries with embeddings. On top of that, a synthesis layer builds 7 cross-cutting architectural narratives (auth flows, DI wiring, data access patterns, microservice communication, etc.) — knowledge that doesn't exist in any single file. The result: when an agent asks "how does authentication work across services?", retrieval returns ~2,000 tokens of already-understood, labeled, architecture-level context — not raw code that the model still needs to comprehend. By your own "smart zone" framing, pre-distillation keeps you there longer because every retrieved token is pure signal. The whole thing runs as a custom MCP server (4 tools: ask_codebase, analyze_story, analyze_bug, get_status) alongside a separate DB MCP with live SQL access to all 16 databases, plus domain-specialized agents for the framework and batch subsystems. Entirely local, zero cloud cost. Your article nails the thesis — context is everything. I'd argue there's a spectrum: Sourcegraph optimizes retrieval precision, this approach optimizes retrieval density. The ideal is probably both. Would love to compare notes. Ajay Sridhar couldn't respond/shortening this - as comment wouldn't allow it. cc: Suresh Kumar Arunachalam, Vivek Chaudhary, Madhan Rangaswamy

🚀 Just open-sourced my end-to-end ML pipeline — fully reproducible, fully tracked! One of the biggest pain points in real-world ML is reproducibility. You train a great model, tweak a parameter, and suddenly can't remember what worked. I built a project to solve exactly that. 🔧 ML Pipeline with DVC + MLflow + DagsHub This project walks through building a production-grade ML workflow where every experiment is tracked, every dataset version is logged, and every pipeline run is reproducible — deterministically, across environments. What's inside: 🗂️ DVC — Versioning data, models & pipeline orchestration via dvc.yaml 📊 MLflow — Logging hyperparameters, metrics & artifacts per run ☁️ DagsHub — Remote storage + hosted MLflow tracking server (no local infra needed!) 🌲 Random Forest Classifier on the Pima Indians Diabetes Dataset Pipeline Stages: ✅ Preprocess → Train → Evaluate (3 clean, modular stages) ✅ Only stale stages re-run on dvc repro — saves time every iteration ✅ Side-by-side experiment comparison via MLflow UI This isn't just a demo — it reflects the kind of MLOps discipline I believe every ML project deserves. Reproducibility shouldn't be an afterthought. 🔗 GitHub: https://lnkd.in/eMWMzcQK 📈 View Experiments: https://lnkd.in/egw6v3di If you're building ML pipelines without DVC + MLflow, give this a look — it might change how you work. Share it if you find it useful! #MachineLearning #MLOps #DVC #MLflow #DagsHub #Python #DataScience #ReproducibleML #OpenSource

🚀 Just built "The Corporate Recon Swarm" — my fastest AI agent orchestration yet! 🏢 What does it do? (The Use Case) — Ever asked an AI to research a company and waited forever while it searches step-by-step? I fixed that. You just feed this Swarm a company name. A "Manager" AI instantly breaks the task down and spawns multiple parallel agents to hunt down their Competitors, Tech Stack, and Recent News at the exact same time. Finally, it merges everything into one master analysis report. ⚙️ How it works under the hood — To pull this off, I moved away from traditional sequential graphs and implemented a Dynamic Map-Reduce (Fan-Out/Fan-In) architecture using LangGraph. 🔹 Dynamic Fan-Out: The Manager doesn't use hardcoded paths. It dynamically spawns concurrent workers using the Send API. 🔹 State Isolation: Each parallel worker runs in its own isolated state. No context pollution, zero token waste. 🔹 Speed & Scale: 10 research queries? It spawns 10 workers instantly. Scaling AI is no longer about just getting an answer; it’s about compute efficiency and orchestration. Project Link : https://lnkd.in/gWu3hbZU #AgenticAI #LangGraph #Python #SystemArchitecture #SoftwareEngineering #BuildInPublic

🚀 Beginning My Journey with LangGraph I’ve recently started working with LangGraph, and it is already reshaping how I approach the design of AI systems. Traditional pipelines often rely on linear chains of execution. In contrast, LangGraph introduces a graph-based architecture that enables the development of stateful, multi-step AI workflows—closely mirroring real-world processes and decision-making systems. 📚 Key Concepts I’ve Learned 🔹 Foundations of LangGraph Understanding how graph-based workflows differ from traditional chaining approaches, and why they are better suited for complex systems. 🔹 Environment Setup & Project Structure Installing dependencies and organizing a scalable development environment. 🔹 Node Creation Designing functional components (nodes) that handle specific tasks within a workflow. 🔹 Flow Control with Edges Connecting nodes and implementing conditional logic to support dynamic execution paths. 🔹 State Management Maintaining and updating shared context across multiple steps to enable more intelligent agent behavior. 🔹 Graph Execution Running workflows while tracking how data flows between nodes. 🔹 Debugging & Observability Improving transparency by tracing execution paths and understanding decision points. 🔹 LLM Integration Combining structured logic with language models to build more capable AI applications. 💡 Key Takeaways Graph-based workflows provide greater flexibility and control Stateful systems enable context-aware AI agents Structured execution improves reliability and maintainability Enhanced observability leads to better debugging and optimization 🛠️ Sample Application As part of my learning process, I developed a sample project demonstrating these concepts: 👉 https://lnkd.in/e3Jf2xtD 🔜 Next Steps Integrate LangGraph into a production-level chatbot system Explore advanced memory handling and tool integration Deploy scalable, real-world AI workflows LangGraph is proving to be a powerful framework for building the next generation of intelligent systems. I look forward to continuing this journey and collaborating with others in the space. #AI #LangGraph #MachineLearning #LLM #Automation #Python #SoftwareEngineering #TechInnovation

Building production-grade AI is about moving from simple RAG to Agentic Workflows. This architecture, built with LangGraph and Pydantic, shifts the focus from linear chains to dynamic orchestration. Key Pillars of the Workflow: The Orchestrator: A central agent that makes real-time decisions, choosing between Knowledge Bases (Vector DBs) and specific Tool Use (SQL/Python). Self-Correction: Unlike basic RAG, this system uses feedback loops. If data is missing, it re-queries until the goal is met. Guardrails & Eval: MLOps in action. Every response is validated for accuracy and safety before reaching the user. Human-in-the-Loop: A critical pause for high-stakes decisions, ensuring trust and reliability. Tech Stack: LangGraph (State management), Pydantic (Type safety), and FastAPI (Production deployment). This is how we bridge the gap between "cool prototype" and "enterprise-ready system." #AI #AgenticSystems #MLOps #LangGraph #GenAI

𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗶𝘀 𝗲𝗮𝘀𝘆. 𝗥𝘂𝗻𝗻𝗶𝗻𝗴 𝘁𝗵𝗲𝗺 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗶𝘀 𝘄𝗵𝗲𝗿𝗲 𝘁𝗲𝗮𝗺𝘀 𝗴𝗲𝘁 𝘀𝘁𝘂𝗰𝗸. After months of community feedback on 𝗠𝗖𝗣 𝗠𝗲𝘀𝗵, we published a 10-chapter tutorial that takes you from a single tool agent to a 13-agent system running on Kubernetes — with multi-LLM failover, distributed tracing, mTLS security, and a mobile-first UI. The key insight behind MCP Mesh: the code you write on Day 1 is the same code running on Kubernetes on Day 9. No rewrites, no sidecars, no platform-specific wiring. Just Python functions with decorators. The mesh handles discovery, dependency injection, health-aware routing, and observability. What the tutorial covers: → Dynamic agent discovery via DDDI (no hardcoded service URLs) → Tag-based LLM routing with automatic failover → Committee of specialists pattern for refined AI outputs → Distributed tracing across multi-agent chains → Production deployment with Helm charts If your team is building multi-agent systems and hitting walls around orchestration, observability, or scaling — MCP Mesh might save you a few months. 𝗧𝘂𝘁𝗼𝗿𝗶𝗮𝗹: https://lnkd.in/e57gWm75 𝗚𝗶𝘁𝗛𝘂𝗯: https://lnkd.in/eqtPvCFZ #AI #MCP #MCPMesh #AgenticAI #Kubernetes #DistributedSystems #OpenSource

Everyone keeps talking about MCP and honestly? My first reaction was that it sounded overly complicated. 😅 Model Context Protocol this. Agentic tooling that. Claude can now connect to your filesystem, your browser, your database... I nodded along in conversations like I knew exactly what was happening. I did not. But here's the thing. Every time I scroll past an MCP demo, I feel that same feeling I got before I dove into RAG systems. That "okay this is actually kind of insane" feeling. And I've learned to trust that feeling. So I've decided to stop nodding and actually start learning. I'm a junior AI Engineer who's spent the last year building RAG pipelines, deploying LLM products, and figuring things out by just jumping in. MCP and LangGraph feel like the next two things I need to understand properly. MCP for connecting AI to the real world. LangGraph for building agents that actually think in steps and don't fall apart on complex tasks. Two rabbit holes. One very curious AI engineer. 😄 Better late than never, right? 🙌 If you've already explored either of these, built something with them, or just have a good resource that helped it click for you, drop it in the comments. I'm all ears. Let's learn together. 👇 #MCP #ModelContextProtocol #LangGraph #GenerativeAI #AIEngineering #LLMs #KeepLearning #AIAgents

Your AI coding agent is only as good as the tools it can access. Paste in some code, get an answer. Paste in more code, get a better answer. But your codebase has thousands of files across dozens of repos... and you can't paste all of it. That's the ceiling most engineers are hitting right now. If your company is on Sourcegraph Enterprise, that ceiling is gone. In case you hadn't heard, we announced an MCP server in our 7.0 release. One command connects it to Claude Code (Anthropic). Once connected, your agent can: - search your entire codebase - navigate cross-repo definitions - pull any file at any revision - trace commit history - invoke Deep Search to answer complex technical questions with sourced answers from your actual code. Here's my write up on how to connect your Claude Code & Sourcegraph accounts together in less than 60 seconds. Link to the writeup in the comments. #claudecode #sourcegraph #mcp #agents #ai

We built an #MCP server for Scrubby to deliver its vast #CodebaseIntelligence context directly to your favorite coding agent. This post looks at why that approach matters and how Scrubby can super-charge any #AI tool you're using in your development workflow. Enjoy! https://lnkd.in/gh7CBjBR #AgenticCoding #ClaudeCode #AICoding #ModelContextProtocol #CodingAssistants #DeveloperExperience

When Liz Pantalone and I first started discussing Scrubby months ago, we imagined the beta as a standalone VSCode plugin, with other IDE or AI API integrations to come a bit down the road. Of course this was the wrong approach 🤦 so we quickly moved to an #MCP layer that would open up Scrubby's #CodebaseIntelligence context and domain knowledge to any AI tool that needed it. TLDR: This was the right decision, lol. 🕺 Go check out this post from our blog about what we considered when designing the server and why it's essential for any serious #AI #CodeReview tool.