Skip to content

Harness → Loop → Graph

How I build agents

I start with the controls around an agent, check its behaviour, then make routing explicit. These open-source reference projects use synthetic data to demonstrate each part of the build.

The counts below are reported test runs from the linked repositories and portfolio evidence.

Harness → GuardrailAI + ProtoBridge

GuardrailAI runs deterministic checks before the model. Its reference API has no authentication or authorization layer, and its hash-chained audit ledger lives in memory. ProtoBridge shows protocol-level tool access through MCP and agent-to-agent requests through A2A. Its bundled tools and Twenty CRM tests use synthetic, offline fixtures.

Loop → evals/cases + per-repo test suites

Run a case, inspect the failure, change the code, and run it again. The portfolio eval cases check which tools and routes the agent used. The eval README records that model quality still needs a run. Each repository also has a suite for testing its code.

Per-repo test suites · REPORTED runs

Graph → OrchestratorX

OrchestratorX uses a supervisor and typed state to choose which specialist runs next. A hop budget bounds execution, and its routing tests run offline.