hebel. today · 8 November 20267 / 73
Agentic Engineering · AI engineering
Building your own agent harness
Core loop mechanics. Paweł Huryn built a 1,200-line Python harness that runs a reasoning loop, tool execution, and session management, which performed in the top group across Claude Haiku and DeepSeek V4.1 Flash models.…
From 7 newsletters · 1 min read · one thing to try
Core loop mechanics. Paweł Huryn built a 1,200-line Python harness that runs a reasoning loop, tool execution, and session management, which performed in the top group across Claude Haiku and DeepSeek V4.1 Flash models (Paweł Huryn from The Product Compass). Sarthak Rastogi structures enterprise agent systems by using a gateway that validates tokens and enforces IAM roles before any tool execution, ensuring that agents only access authorized resources (Sarthak Rastogi from AI Agent Engineering).
Environment simulation. Raindrop reconstructs production environments using stateful synthetic services, allowing developers to test how model swaps or prompt edits affect agent behavior in isolation (The New Stack). The WorkFactory project verified 112,686 tasks across 612 occupations by checking if simulated workspaces contain the information an agent needs to complete a job (🔳 Turing Post).
Skill-based performance. NVIDIA increased average correctness scores on specialized tasks from 46 to 87 out of 100 by implementing skills that provide agents with concrete procedures and domain expertise (Salvador Carranza from What AI makes PossibNow). Researchers testing SWE-Skills-Bench found that 92 of 98 tested skills produced no statistically significant effect, indicating that only skills supplying specific, actionable procedures improve agent output.
- Spawn's ten-person team shipped 77,000 changes in 30 days by building AI judges from two years of recorded decision data (Web Development Newsletter by Packt).
- WebAnything reconstructed 244 websites to train agents, raising Qwen3.5-9B’s success rate by 13.33 percentage points (🔳 Turing Post).
Try today · 40 min
Build a basic reasoning loop for your agent that executes a single tool call and logs the result.
Start by writing a loop that sends a prompt to the Anthropic API and prints the tool call request before executing it.
You need
- Anthropic API key
- Python environment
- A simple tool function (e.g., file reader)
Steps
- Initialize a conversation list with a system prompt.
- Send the conversation to the model with tool definitions.
- Parse the model's response for a tool call request.
- Execute the requested tool locally.
- Append the tool output to the conversation history.
- Send the updated history back to the model for the final answer.