Speak at The Fifth Elephant 2026 Annual Conference

Abstract

Everyone can get an AI agent working in a demo. Keeping it working - and understanding why it behaves the way it does, is a different skill. And it turns out that skill has surprisingly little to do with the model. It’s about the harness: everything you build around the model to turn it from something that talks into a system that acts.

In this hands-on workshop, we build a minimal coding agent from scratch - a tiny Claude Code - and use it as a lens to understand how agentic systems really work. How harness engineering is done. We keep the build small and focus on a few core primitives: an agent loop, a couple of tools, and context engineering - including skill support, mcp support and compaction, so you can understand exactly what happens inside the context window as an agent runs. Around that build, we map the whole landscape: how today’s agents (Claude Code, Codex, Cursor, OpenClaw, Pi, Deep Agents, and more) are actually put together, and the concepts that define the frontier - memory, compaction, streaming, MCP, skills, dreaming etc.

Key takeaways

The working minimal agent is a means, not the end. By the end of this workshop, participants will leave with:

• First principles of how to build a harness - the core primitives every agentic system is made of, and how they fit together.
• A mental model for reasoning about any agentic system - how to think about agent behaviour, reliability, and cost in terms of what’s happening inside the context window.
• A clear map of the landscape today - the tools, the patterns, and the frontier concepts (memory, compaction, dreaming, streaming, MCP, skills) and how to reason about each.
• A working minimal coding agent - a tiny Claude Code they built and understand end to end, to keep experimenting with.

Who this workshop is for

This workshop is designed to be beginner-friendly. You’ll get the most from it with a basic understanding of LLMs, some familiarity with agentic frameworks, and a working knowledge of Python. Having built an agent before, even just a demo, is a plus, but it is not required, and no prior AI/ML expertise is needed.

Speaker

Shubhankar is a Senior Software Engineer 2 at Atlan, where he heads internal AI for the company. A generalist who’s gone deep on agents, he builds agent platforms at scale that handle real work across the company - from the software development lifecycle to finance, legal, and CX. He’s been hands-on with agents, day in and day out.

Past talks & workshops with Fifth Elephant

Build MCPs: Make Your App AI-Controllable — hands-on workshop, The Fifth Elephant 2025. https://hasgeek.com/fifthelephant/mcp-hands-on-workshop/

Outline

The session moves in stages rather than on a strict clock, mixing short concept segments with hands-on building. Roughly 40% is discussion (the need, the landscape, the frontier) and 50% is building (the loop, tools, compaction).

• Why harnesses matter - the need. Why a demo works and production doesn’t, and why that gap isn’t about the model. The frontier models have largely converged, so the harness around the model now decides the outcome.

• The landscape. How agents are actually built today - terminal agents, IDE agents, cloud/async agents, and managed-agent platforms - across tools like Claude Code, Codex, Cursor, OpenClaw, Pi, Deep Agents, and OpenCode, and the shared standards that tie them together.

• First principles. The one idea everything else hangs on: Agent = Model + Harness. The primitives every agentic system is built from, and why almost every decision comes down to what’s in the context window right now.

• Build: the agent loop (hands-on). The heart of every agent - model, tool calls, execution, repeat. We build it so it supports sub-agents out of the box.

• Tools (hands-on). Giving the model hands: reading files, editing files, running commands. The point where a model that could only talk starts to act.

• Context engineering ( hands-on). The context window as the agent’s working memory, and how the harness curates it. This is where Skills and MCP fit, and where we cover the techniques for keeping the window healthy as a task grows: compaction

• Build: compaction (hands-on). We implement compaction directly, so you can see the context window stay manageable as the agent runs - making the abstract concept concrete.

• The frontier - discussed, not built. The popular harness concepts worth knowing even though we won’t build them all: long-term memory, “dreaming” (offline memory consolidation — Claude’s Dreams, OpenClaw’s sleep cycles), streaming, and the rest of the advanced toolkit.

• Wrap-up & discussion. Recap the mental model, the one thing a harness can’t do for you (write the spec), and open Q&A.