The Hypothesis

If an agent has all the skills, tools, and prompts it needs — what's the smallest prompt that gets it running fully autonomously?

The Concept

Context is expensive. Every token in a prompt costs latency, money, and attention. If the agent already has access to a full registry of skills, tools, and prompt templates — via something like Orchesta — how little do you actually need to say to get it started? This experiment strips the initial prompt down to the absolute minimum while still achieving fully autonomous, zero-human-intervention execution.

Minimum viable prompt

The hypothesis

If an agent has all the skills, tools, and prompts it needs — what’s the smallest prompt that gets it running fully autonomously?

The concept

Context is expensive. Every token in a prompt costs latency, money, and attention. If the agent already has access to a full registry of skills, tools, and prompt templates — via something like Orchesta — how little do you actually need to say to get it started? This experiment strips the initial prompt down to the absolute minimum while still achieving fully autonomous, zero-human-intervention execution.

How it works

1. Full prompt baseline — run a task with a detailed, hand-crafted prompt
2. Strip to essentials — remove context the agent can look up itself
3. Test autonomy — does the agent still complete the task without intervention?
4. Measure quality delta — compare output quality vs. full prompt
5. Find the floor — keep compressing until the agent breaks or drifts

Progressively shrinking the prompt until you find the minimum viable instruction for full autonomy.

What it explores

• What’s the minimum context an agent needs to start a task if it has access to a tool/skill registry?
• Does a leaner prompt produce faster execution with no quality loss?
• At what point does prompt compression cause the agent to drift or hallucinate?
• Can the agent learn to fetch its own context instead of being given it upfront?
• How much of a typical prompt is redundant when the agent has access to Orchesta?

What we found

• Prompts could be compressed by 60–70% without measurable quality loss when the agent had access to a skill registry
• The agent successfully fetched its own context for tool schemas, style guides, and task templates
• A one-sentence prompt (“Run the email campaign for segment X using the Q4 playbook”) worked when the agent could look up everything else
• Vague prompts (“do the email thing”) caused drift even with full registry access

Learnings

• Give the agent a pointer to context, not the context itself — knowing that a registry exists is more valuable than pasting its contents into the prompt.

• The minimum viable prompt has two requirements: clear intent and clear scope — everything else can be fetched.

• Fewer input tokens means faster time-to-first-action — leaner prompts are a latency optimisation, not just a cost one.

• Vague intent is the failure mode, not missing context — invest prompt tokens in specifying what you want, not in background information the agent can look up.

Where it goes next

This has direct implications for how Orchesta’s skill registry is designed. If agents can fetch their own context, the registry needs to be optimised for discoverability — not just storage. We’re exploring whether a search-first interface (like Quiver’s approach) works better than a structured catalogue.