Reconstructing the agent methodology: Decoupling decision-making and execution - open source [P]

I’ve been thinking about a problem in current agent systems:

Most agents are becoming very good at execution, but the decision layer before execution is still unclear.

Coding agents, research agents, tool loops, sandboxes, workflows, and harnesses are all improving quickly. Once a human gives an intent, agents can often do a lot of useful work.

But the higher-level question is still usually left to the user:

What should happen next, and why?

I’ve been exploring this idea through an open-source project called Spice.

The simplest way to describe it is:

Spice is a decision layer above agents.

It is not trying to replace execution agents. Tools like Claude Code, Codex, Hermes, or other agents can still do the actual work.

Instead, Spice sits before execution and tries to make the decision process explicit:

• what was observed
• what options were considered
• why one option was selected
• what trade-offs were rejected
• whether execution needs approval
• what happened afterward
• how that outcome should affect the next decision

The current runtime is still early, but it can already be installed, configured with an LLM provider, run in the terminal, inspect Decision Cards, and hand off approved execution to external agents.

The goal is to make agent behavior less of a black box.

Instead of only seeing the final result of an agent task, I want to preserve the reasoning boundary before execution: what the system believed, what it chose, why it chose it, and what changed after the action.

GitHub: https://github.com/Dyalwayshappy/Spice

I’d love feedback from people building agents. Feel free to fork, star the repo, or share any feedback and ideas. Would love to build this together with the community.