Building an Internal Platform for AI Agents
A technical guide to the control plane behind production coding agents: workload selection, identity, sandboxes, context, workflow graphs, CI, observability, and human control.
Siddhant Khare
Software engineer on Okta's Emerging Technology team and former early engineer at Ona (formerly Gitpod), which OpenAI has agreed to acquire for Codex. Writing and building around review burden, context discipline, agent observability, and governable AI-assisted workflows. OpenFGA maintainer. Author of The Agentic Engineering Guide.
Latest posts
The review queue is the bottleneck
Agent coding tools need to make their work reviewable, not just generate more code.
Retry is not a loop, its a data structure
Under partial failure, retry loops can rearrange history. If order, duplication, or staleness matters, retry needs state.
Your agent should not inherit your admin token
Agents need delegated access, not inherited human access. Put policy between the agent and the tool before agents touch real systems.
My first company is joining OpenAI. I started there at 19.
What four years at Gitpod and Ona taught me about public work, trust, messy ownership, and getting closer to rare work.
Build your own AI experimentation stack
Netflix spent years building their experiment platform. You have the tools and a weekend. The hard part is not spinning things up. It is tearing them down cleanly when the idea fails.
What's next for agent-trace
agent-trace shipped two months ago as a debugging tool. People started asking how to run it in production. v0.4.0 and v0.5.0 are the answer.
Bangalore is not an upgrade
I turned down two roles at global tech companies this year. Not because of the money. Because of relocation. The assumption global companies keep getting wrong about hiring in India.
I built an AI agent that orders food for your entire team
Your team fills a Google Form. The agent reads it, picks a restaurant, builds a Swiggy cart per person, and waits for you to confirm.
Attackers are now targeting your AI coding tool
The Bitwarden CLI supply chain attack explicitly hunted Claude Code, Cursor, and Codex CLI configs. The same week, DPRK confirmed slopsquatting works in the wild. Here is what happened and what to do.
The Vercel breach started at a tool nobody was watching
The Vercel breach did not start at Vercel. It started at Context.ai. And the fix is not rotating credentials. It is a different credential model for AI tools entirely.
Why you stop caring mid-review
There's a moment in every long review cycle where you quietly check out. It doesn't look like giving up. It looks like being collaborative.
Tracing what agents do, not what they say
Stripe merges 1,000+ agent PRs per week. Spotify vetoes 25% of agent sessions. Cursor runs hundreds of agents in parallel. Dashboards show the model layer. Traces show the system layer. You need both.
The plumbing behind Claude Code
I read Claude Code's leaked source. Not magic, just good engineering. 10 patterns that show what actually matters when building AI tools.
AI Agent stack you need Context, Auth, and Cognitive Debt
Most AI content teaches you how to write prompts. This is not that. I've spent three years at Ona...
The ambition tax
Nobody told me that ambition has a tax. Not the long hours or the missed weekends. The quieter kind.
The agent observability gap
We have better observability for a Node.js microservice than for an agent that just rewrote half a codebase. Here is what is missing.
Claude Code's broken permission model
Fifty approval prompts or full shell access. Nothing in between. The permission model needs infrastructure, not prompts.
You're tired because your AI has no feedback loop
Not prompts. Not models. Feedback loops. The engineering discipline that separates teams thriving with AI from teams drowning in reviews is called backpressure.
AI fatigue is real and nobody talks about it
You're using AI to be more productive. So why are you more exhausted than ever? The paradox every engineer needs to confront.
Ona (formerly Gitpod) is re-launching its Open Source program
Gitpod started as an open-source project. Long before “AI productivity” became a thing, the core...
Containers aren’t a sandbox for AI agents
Where containers stop being simple Containers are sold as a solved abstraction. You...
2025 was the year everything changed
Bought my first home. Gave my first international conference talk. Shipped AI tools used by Fortune 500 companies. My honest 2025 wrap.
The Engineering guide to Context window efficiency
A deep dive into semantic deduplication for LLM context windows If you're building with RAG...
Beyond finding: Remediating CVE-2025-55182 across hundreds of repositories with Ona Automations
Finding vulnerable code is only half the battle. When a critical CVE drops, engineering teams face a...
Securing Agentic AI: authorization patterns for autonomous systems
Why traditional access control fails for AI agents, and how relationship-based authorization provides...
Context Engineering: The critical Infrastructure challenge in production LLM systems
The $10M question nobody's asking While the industry obsesses over model parameters and...
AWS S3 Vectors at scale: Real performance numbers at 10 million Vectors
Introduction AWS S3 Vectors promises "billions of vectors with sub-second queries" and up...
Why agent orchestration is harder than kubernetes - Lessons while building Agentflow
TL;DR: While building AgentFlow, an open source orchestration engine for AI agents, I discovered...
Serverless economics: why Cloud Run crushes App Runner (until it doesn’t)
This analysis is based on official pricing documentation and straightforward cost calculations. ...
What I'm trying to escape as an Indian engineer
Competence earns respect, but originality creates legacy.
Caring pays off
Whatever you do, care. Care enough to notice. Care enough to act. Care enough to finish things well.
Ready-to-eat knowledge economy
Your AI teacher is free, available 24/7, and infinitely patient. Stop asking for permission to learn.
The visibility game
Organizations will always reward perceived value over hidden value. The question is whether you'll play the game as it is.
GPU memory virtualization for LLMs
LLM inference is shifting from compute-bound to memory-bound. Treating GPU memory like an OS does RAM is the path forward.
Attention by subtraction
How to make AI code edits more accurate
A technical examination of production-grade LSP-MCP integration After spending months analyzing AI...
The work finds you back
Ship. Share. Repeat.
Most people think you build skills, then market them. It turns out this is backwards. The market is always watching, and the feedback loop is so long you don't notice.
An easy way to stop Claude code from forgetting the rules
You spend time setting up Claude Code with specific instructions in your CLAUDE.md file. Maybe you...
Why context engineering matters more than prompt engineering
Prompt engineering optimizes the question. Context engineering optimizes the information the model sees before it answers. Here's why the latter wins.
Claude code is costly - unless you do this
Claude Code is a powerful AI development tool, but it comes with a rate limit if you're on the Pro...
Authorization patterns for AI agents: beyond API keys
API keys give all-or-nothing access. AI agents need fine-grained, task-scoped, auditable authorization. Here's how to build it with OpenFGA.
Extending Claude's Brain: How MCP makes AI actually useful for developers
Model Context Protocol (MCP) is a protocol created by Anthropic designed to enable AI models to...
How docs AI search works: Mintlify-Style with OpenAI Agents SDK
Documentation remains a critical resource for developers, serving as the definitive guide and...
Surviving tech bureaucracy: How genuine competence cuts through buzzwords
Isolating AI Agents with DevContainer: A secure and scalable approach
AI coding agents like Cline and RooCode are powerful but unpredictable. A simple misconfiguration...
Tackling Cold Starts in AWS Lambda: A Deep Dive with LLRT
AWS Lambda is the go-to serverless computing service for developers who love event-driven...
Engineering with intent: A mental checklist for clear thinking
How we slashed CI build time using Go’s cache
Our CI pipeline’s integration tests were running sluggishly. The primary bottleneck wasn't the tests...