Deep Dive into Agents

The word agent gets used very loosely — sometimes it means a chatbot with a tool attached, sometimes a long-running autonomous system. This talk uses a narrower engineering definition and then spends its time on the part nobody puts on a leaderboard: the harness, the software wrapped around the model call.

The claim in one line: two agents running the same model weights can differ by seven to ten benchmark points, and the difference is entirely the harness. If that holds, choosing an agent is not a model-selection problem — it’s a systems-engineering one.

Present from the fullscreen link — the deck is a standalone page with no site chrome. Each slide is addressable by hash, so /agents-deck/#17 opens directly on the Claude Code section, and the position survives a reload.

What the deck covers

First principles. What separates an agent from a chatbot or a script: who selects the next action. Then the loop every agent here reduces to — call the model, execute its tool calls, append results, repeat — and the long list of things that loop deliberately does not specify.

Harness anatomy. Seven slides on the design axes: one tool for everything versus one tool per job, and what each access model costs; the context window as working memory; why a permission prompt and a sandbox solve different problems; how hooks, skills, MCP and plugins do four different jobs; what a saved session buys you; and where the loop actually runs.

Five implementations, five bets.

• Claude Code — bet on the model. A single-threaded decision loop with parallel tool execution, tools that refuse rather than guess, scoped memory files riding the cached prefix, and mode → rules → sandbox layering.
• pi — start small. Four active tools, a fixed prompt under 1,000 tokens, JSONL session trees, MCP deliberately left out of core so its context cost stays opt-in.
• opencode — the agent is a server. Many clients, 75+ providers, and configuration as the primary product surface.
• DeepSeek Harness — even the loop is a plugin. Event-driven execution with an append-only log, and four selectable run modes.
• Hermes — the agent that remembers. A learning layer of memory, user modeling and reusable skills wrapped around the execution loop.

My take. The mechanics converge; the beliefs don’t. Each project answers a different question about where intelligence, control and durable state should live.

Three things to hold loosely

The deck says this up front and it’s worth repeating here:

1. It’s a snapshot dated August 2026. These projects ship weekly — flags, defaults and whole features change. The architectural ideas are more durable than the exact counts.
2. Claude Code has a closed core. Some internals come from a leaked build rather than published source, and Anthropic hasn’t confirmed them.
3. The numbers aren’t a leaderboard. The referenced eval is 30 tasks with one model, and not every harness ran on the same settings. Stars measure attention, not quality.

Sources are cited on each slide.