Agent Runner

Define, deploy, observe, and govern autonomous agents

Agent Runner is a flexible, configuration‑driven framework for building and running AI agents equipped with tools, memory, and custom actions—all declaratively defined in YAML or JSON. Spin agents up from a shared runtime instead of writing bespoke glue code every time.

The Problem

When I set out to build my first agent, I realized the real bottleneck wasn’t the agent itself—it was the plumbing required to run any agent. I already had a dozen ideas; the overhead of spinning up infrastructure for each would kill momentum.

Lessons learned as an analyst, coder, and team lead: don’t build one‑offs. Build tools, frameworks, and productized workflows.

So I built Agent Runner—a runtime that loads, executes, and monitors agents from a simple YAML spec. Under the hood it leverages OpenAI’s Agents SDK, providing trace logging, tool registries, lifecycle hooks, and real‑time feedback—all Python‑native and pluggable.

Core Capabilities

Declarative configs for tools, memory, behaviour, & flows
Extensible action registry—any Python function becomes an agent action
Dynamic tool loading & resolution
Optional short‑term or persistent memory per agent
Output writers: JSON, Markdown, or plain text
CLI runner for one‑command execution
Tight integration with OpenAI Assistants API
Guardrails for safe, bounded behaviour
Interactive chat mode for live debugging

Architecture Overview

AgentRunner/
├── actions/        # Custom functions
├── agents_modules/ # Agent loading & config
├── common/         # Shared infrastructure
├── config/         # YAML/JSON specs
├── main.py         # CLI entrypoint
├── scripts/        # Utilities
├── tools/          # Tool implementations
├── memory/         # Memory back‑ends
├── outputs/        # Writers & logs
│   ├── agent_run_writer.py
│   ├── usage_stats_writer.py
│   └── ...
└── inputs/         # Example configs & data

Use‑Case Snapshot: Cleveland Cup Reporter

Query latest golf‑league scores via SQL
Identify factors behind top finishers
Generate a concise event recap for the league website