I build agents, and the MCP servers, CLIs and SDKs they run on. AI engineer, 8+ years.

himadri.dev · Agent Experience field guide · LinkedIn · X

At Mudita Studios, where I build AI products end to end (private work)

• A coding agent that takes a Slack or Jira request to a tested, reviewed draft pull request. It works in a sandbox, tests its change in a real browser, and a second model checks the result before a person reviews it.
• A research product in production, where every claim in a report cites a source or is marked unverifiable, built on a sandboxed agent runtime with spend limits and tracing.

At Firecrawl, where I owned the Agent Experience programme (merged open-source PRs)

• Measured and improved how AI agents discover, choose and correctly use Firecrawl across Claude Code, Codex, Cursor and other agents, and shipped the fixes.
• Contributed to the hosted MCP server: OAuth and keyless setup, a move onto upstream FastMCP, and errors that tell an agent how to recover.
• Improved tool descriptions, docs, CLI help and skills so agents choose the right tool, and added CI checks that run the docs' code samples.
• Built a daily benchmark of whether agents pick a tool and use it correctly, and a sandboxed experiment harness behind the shipped fixes. Agent Experience owner for the Developer Index and Government and Legal launches.

My own work

• agentexperience.tech: a field guide to Agent Experience, with an agent-readiness rubric and a public, read-only MCP server.
• awesome-agent-experience: a curated library of sources on tool use, discovery and evaluation.
• qwen-3.6-35b-consumer-gpu: Qwen3.6-35B running at 43 tokens per second with 128K context on my 8 GB laptop GPU, with launch scripts, a tuning guide and a coding benchmark.

• A tool call is not a finished task. I check the result, not just whether the answer sounds right.
• Agents read text, not screens. Tool descriptions, error messages and docs are the interface, so I write them for the agent.
• A person stays in the loop wherever work ships.

Before that, I built Knit's agentic research platform, which cut report turnaround from 2-3 days to under an hour, after nearly six years of production ML in search, recommendations and computer vision.

Some professional work is private. I share the ideas and patterns, not client code, names or internal results.