Agentice Stack: A Model-Agnostic AI Agent Harness, Open Sourced

We just open-sourced the agent harness we use for every deployment: Agentice Stack. MIT-licensed, two dependencies, and it runs fully offline for demos.

https://github.com/atlas-agent922/agentice-stack

Why a harness, not a framework

Most agent frameworks sell you the model. The model is the commodity — GPT-4 today, something else next quarter, and a local 8B model when your compliance team asks "where does this data go?"

The harness is the asset. It's the loop, the tool registry, the guardrails, and the trace log — the code that survives when you swap models. If your agents are built directly on one provider's SDK, you don't have an agent system, you have a vendor relationship.

The design rules

• The model is a config value. provider: openai today, provider: ollama on your own GPU next month. No code changes.
• Guardrails are checked before every step. Iteration cap, token budget, cost circuit breaker, repetition detection, and approval gates on sensitive tools. An agent that can't be stopped isn't an agent, it's a liability.
• Every decision is traced. JSONL log, one line per step: the prompt, the decision, the tool call, the result, the tokens, the cost. When something goes wrong, you replay the exact chain. You don't guess.
• Max 7 tools per run. Tool overload measurably degrades LLM tool selection. Capping the registry is a reliability feature, not a limitation.

What's in the box

agentice/
├── config.py       # one YAML file = the whole system
├── llm.py          # OpenAI-compatible + offline stub backends
├── tools.py        # tool registry with a 7-tool cap
├── guardrails.py   # iteration / token / cost / repetition / approval
├── agent.py        # the loop, ~150 lines, readable in a sitting
├── trace.py        # JSONL decision log with replay
└── demo.py         # runnable demo, works with no network

Eight unit tests cover the happy path and every failure mode: iteration cap, cost breaker, repetition guard, approval gate, tool-cap enforcement, and trace replay.

The point

If you're evaluating AI agents for a business that can't send its data to a third-party API, this is what the foundation layer looks like. Small enough to audit in an afternoon, boring enough to run unattended.

We'll be building in public on top of this — the next post is the first production teardown from a real deployment.