Staff Software Engineer, AI Platform @ Cendyn

Hi, I'm Akshit

Building production-grade AI infrastructure: LLM gateway, MCP infrastructure, and agent harnesses. Nine years from frontend architecture to tech lead to Staff, so interfaces get the same care as infrastructure. I write about what breaks.

01 From the garage

Open the garage

pyforjs

Python for senior JS/TS engineers: learn 3.14+ through TypeScript→Python translation, capped with an AI agent built on CPython running in-browser via WASM.

pi-agent-proxyTypeScript · Pi Extension · ObservabilityNative observer for pi. Token audits, tool inventory, and wire transcripts with no HTTP proxy.

external-model Agent SkillTypeScript · Agent Skills · Claude CodeDelegate subtasks to cheaper, faster models like DeepSeek, Qwen, and Kimi from Claude Code and other CLIs.

WebMCP & Agent-Ready WebWebMCP · Browser Tools · StandardsTool manifests served from this site. Machine readable APIs for agents, over web standards.

02 Writing

Browse posts

Engineering Agent HarnessesLatest

Anatomy of an Agent Harness: The Code That Runs When You Type a Prompt

23/08/20265 min readAIEngineering

Hand-drawn illustration of the portling cranking the harness loop between model and execute boxes

The LLM never touches your filesystem. A look inside the agent harness, the run loop that turns model output into actions, and why understanding it makes you better at working with agents.

Agent Harness Guardrails: Permissions, Sandboxes, and Hooks

23/08/2026

A confused model is a security incident waiting to happen. How agent harnesses use permission gates, OS-level sandboxes, and deterministic hooks to keep autonomous work from becoming autonomous damage.

Context Engineering: The Hardest Problem in Agent Harness Design

23/08/2026

Your agent's context window is its entire universe, and it fills up faster than you think. How harnesses fight back with compaction, structured memory, and ruthless curation.

Long-Running Agents: Checkpoints, Resume, and Steering

23/08/2026

Ten-hour agent runs die at hour nine for boring reasons. How harnesses use checkpoints, session persistence, and mid-run steering so long autonomous work survives contact with reality.

03 Reading

Browse books

2026

System Design for the LLM Era

Sampriti Mitra

2026

Atomic Habits

James Clear

04 Daily drivers

Open the stack

pi: Daily harness
Herdr: AI & Agents
MacBook Pro 14-inch (M5): 16GB