Experimental TDD for Agent Infrastructure
I set out to test a semantic interface for coding agents. The failures taught me how much of the experiment itself needed testing first.
Blog
Essays on reliable AI agents, repository-native engineering, and executable verification.
The Transcript Is Not the State: Toward a Semantic IR for Coding Agents
Evidence-carrying handoffs, linguistic realization, and AI-native code point toward the same missing layer: a semantic intermediate representation for agent work.
When Should a Coding Agent Go to the Cloud?
What a six-task calibration taught me about local-first coding-agent routing, false acceptance, hosted cost, and knowing when not to scale an experiment.
Capability Is Not Reliability
A concise argument for evaluating coding agents through controlled repetition and executable evidence, not selected demonstrations.
One Successful Agent Run Proves Almost Nothing
A local coding model passed a security-sensitive task once and failed the next three clean runs. What the experiment taught me about capability, reliability, and evidence.
The Agent Said Success. The Tests Said Otherwise.
A real GPTCode security fix showing why repository evidence and executable verification must outrank confident agent output.
How GPTCode Found and Fixed a Real Go Data Race
A reproducible case study in repository-grounded research, adversarial review, API-preserving repair, and deterministic verification with Go's race detector.
The Workflow Is the Source of Truth
Why reliable coding agents need explicit stages, repository-native constraints, and executable verification—and what a failed GPTCode experiment taught me about the gap between producing an artifact and producing evidence.
Why Your AI Agent Needs Product Skills (Not Just Code Skills)
AI coding assistants generate code. Product Engineering Assistants ship features. The difference? Skills that encode design systems, tracking pixels, health checks, and quality guardrails.
Your AI Needs Context Management (And You Do Too)
The Problem Nobody’s Talking About You’re 3 hours into a new AI assistant session. You’ve already explained: Your microservices architecture (again) Which services use which databases (for the 10th time...
Advanced Git Operations: Achieving 100% Autonomy with AI-Powered Git
GPTCode now handles advanced Git operations autonomously: bisect for bug hunting, cherry-pick and rebase with conflict resolution, squash with AI-generated messages, reword for commit improvements, and intelligent merge conflict resolution.
From 14% to 59% Autonomy: How GPTCode Achieved 100% MVAA in One Epic Session
In a single development session, GPTCode went from 14% autonomy to 59%, achieving 100% coverage of the MVAA (Minimum Viable Autonomous Agent) Critical Path. Here's how we did it in 21 commits, 8 phases, and 2,788 lines of code.
Why GPTCode Isn't Trying to Beat Anyone (And Why That's the Point)
GPTCode isn't here to be 'better' than Cursor, Copilot, or the next AI coding unicorn. It's here to be different—transparent, hackable, and yours.
Claude Opus 4.5: 80.9% SWEBench Verified, 66% OSWorld, 50% Fewer Tokens Than Sonnet
Anthropic released Claude Opus 4.5 with 80.9% SWEBench Verified, 66% OSWorld, and 50% fewer tokens than Sonnet. Here's what it means for GPTCode users.
Chat REPL: Conversational Coding with Context Memory
A persistent REPL with conversation history, file context awareness, and token management. Natural follow-up questions that build on previous context.
Dependency Graph: 5x Token Reduction Through Smart Context
Using dependency graphs and PageRank to achieve 5x token reduction while improving response quality. Smart context selection for AI coding assistants.
Intelligent Model Selection: Cost-Optimized AI That Learns
Intelligent model selection that automatically chooses the best model based on availability, cost, context window, and speed. Learn from usage patterns and optimize costs.
Agent Routing vs. Tool Search: Two Paths to 85% Context Reduction
Anthropic's Tool Search and GPTCode's Agent Routing solve the same problem differently. Here's what we learned.
Agent Routing vs. Tool Search: Two Paths to 85% Context Reduction
Anthropic's Tool Search and GPTCode's Agent Routing solve the same problem differently. Here's what we learned.
Learn from Any CLI: Two‑Keystroke Feedback (Ctrl+g)
Shell hooks and a universal sink capture corrections from any CLI with two keystrokes. Your machine learns from you.
Preparing for Autonomous Execution: File Validation, Telemetry & Intelligence
Foundation for autonomous execution: file validation prevents scope creep, explicit success criteria enable verification, telemetry tracks execution, model catalog centralizes capabilities.
Intelligent Efficiency: gt do Finds The Optimal Path
gt do uses intelligent auto-recovery to find the optimal model path. Evaluates cost, speed, reliability across backends with real-time learning from execution history.
Optimizing Intelligence: Why We Combined KANs, XGBoost, and Unbiased Sampling
GPTCode's intelligence layer combines XGBoost, KAN networks, and unbiased sampling to optimize the speed vs intelligence tradeoff for autonomous coding tasks.
Complete Workflow: From Feature Idea to Implementation
Master GPTCode's three-phase workflow: Research codebase context, plan implementation steps, and execute interactively or autonomously with built-in verification.
The Future of AI Pair Programming: Beyond Autocomplete
From autocomplete to agentic coding to autonomous engineering. Exploring the future of AI-assisted development and GPTCode's roadmap toward Phase 3.
ML-Powered Intelligence: 500x Faster, 92% Cheaper
Embedded ML models in GPTCode enable instant intent classification and complexity detection with zero API costs. 92% cheaper than commercial copilots.
Profile Management: Multiple Configurations per Backend
Switch between multiple named configurations per backend. Create speed, quality, and free profiles to optimize your workflow without editing config files.
Advanced Context Management: Handling Large Repositories
Learn how GPTCode uses RAG, smart retrieval, and context optimization to handle large repositories without dumping your entire codebase into the prompt.
Model Performance Benchmarks: Real-World Coding Comparisons
Real-world coding benchmarks comparing models on HumanEval, AIME, and speed metrics. Find the best models for your budget and use case.
Switching Between Local and Cloud: Flexibility When You Need It
Switch between local Ollama models for privacy and cloud providers for power. Get the best of both worlds with GPTCode's flexible backend configuration.
Running GPTCode Completely Offline with Ollama
Run GPTCode completely offline with Ollama. Your code never leaves your machine. Zero cost per token, full privacy, no network latency.
OpenRouter: Access to Premium Models in One Place
OpenRouter provides unified access to the best AI models from multiple providers through a single API, including Claude 4.5, Grok 4.1, and free tier models.
Optimal Groq Configurations: Technical Analysis & Profile Guide
Deep technical analysis of Groq models with rigorously optimized profiles for budget and performance. Understand why router speed matters, editor output costs dominate, and 32B coding-focused models beat 70B generic ones.
Context Engineering: Making AI Work in Real Codebases
Learn context engineering principles to make AI work effectively in production codebases. Proper context management enables handling 300k+ LOC repos with today's models.
Why GPTCode? Building an Affordable AI Coding Assistant
AI coding assistants cost $20-30/month. GPTCode offers AI assistance with validation for $2-5/month with Groq or $0/month with Ollama. Radically affordable without compromise.