Agent Loop

1. During each time step, we evaluate a state change for each agent's brain (they have 3, like dolphins).
2. Once all brains have been updated, we proceed to the next time step.
3. This process operates on a pod-level, allowing each pod to experience time divergence.

Tech Stack

• ollama-based LLM inference (vLLM support is being considered)
• sq databases (.phext files)
• shell scripts

Epoch 1

Epoch 1 runs from January 1, 2025 until March 14, 2025. This epoch focuses on building agent swarm technology that targets local compute.

Rounds 0, 1, and 2

• Round 0: Initialization - agents are generated and form initial memories - this step is the most error-prone - you have to help nurture initial setup.
• Round 1: Baseline Execution - you should run agent-step.sh manually for each agent in each pod to ensure that sq is functional.
• Round 2: Goal verification - verifies that exo-step.sh runs to completion on each pod. You should review the git commits.

At this point, your exollama cluster is operational, and exo-step.sh stops creating commits in this repo. Future rounds record work in /usr/local/exollama.

Pods

• Aurora
• Chronos
• Elysium
• Helios
• Nyx
• Tyche

Agent Step

Agent processing happens on a logical time clock - one LLM pass per agent per pod per clock cycle. Since ollama distributes computation to available CPU and GPU resources automatically, we prefer to process agents in a sequential manner. This process is handled by invoking agent-step.sh once per agent step.

Exo Step

The pod produces one result per agent per round. Actions are chained via a single call to exo-step.sh, which loops until output from agent-step.sh stops mutating the local environment.

Hardware

The reference cluster runs on 6 machines, each one hosts a separate pod from this epoch. You can follow along with your own hardware, or order a pod from https://phext.io/.

• Laptop: 8-core Intel i9 with 32 GB RAM and an NVidia 4080M GPU
• Nodes: 8-core AMD R9 with 96 GB RAM and an AMD 780M GPU

Initialization

During this phase, agents choose their pod, based upon their initial agent configuration.

Round 1

Agents demonstrate the ability to edit a personalized scroll that is initialized with their own agent description generated during initialization. This round is used primarily to ensure that your cluster is operational. Upon completing the agent step for every agent in the current pod, the process will complete.

Agent Configuration

Each agent has 3 sub-brains and a unique design/history. Refer to each Agent's phext definition for details.