Adaptive compute
Use compute in proportion to the complexity and needs of the task.
Phylogeny AI / ProjectsExplore Morph
Project 001 / Open research
MORPH
An agent that can change how it thinks.
Morph is an experimental runtime for dynamic collaboration between specialized models and capabilities. It is a research project from Phylogeny AI, a community AI lab.
TYPE Open-source research
FOCUS Dynamic agent runtime
PHASE Early experimentation
01 / The question
Intelligence should be a configuration, not a fixed model.
Most agents start with one language model and attach tools to it. Morph explores a different idea: intelligence is a population of specialized capabilities that a runtime can activate, put to sleep, replace, and coordinate as a task evolves.
An agent does not have one fixed intelligence. It can change its cognitive configuration to fit the work.
02 / Runtime loop
Capabilities follow the task.
After each step, the orchestrator can reconfigure the active population for what comes next.
INPUTAnalyze this sample and explain the anomaly.
01 / REASONINGACTIVE
02 / CODINGSLEEPING
03 / VISIONOFF
STEP RESULTNew evidence found / activate vision
RECONFIGURE POPULATIONREASONING / SLEEPINGVISION / ACTIVE
OUTPUTVerified result + shared context
03 / Lifecycle
Every capability has a state.
The runtime discovers available capabilities and manages their transition through the task.
OFF
No resources loaded.
LOADING
Model weights are being prepared.
ACTIVE
Ready for inference.
SLEEPING
Inactive, resources retained when useful.
UNLOADED
Weights removed to free memory.
04 / Research directions
What Morph is designed to explore.
Resource-aware
Sleep or unload idle local models to return RAM and VRAM to the system.
Capability-first
Depend on what a model can do, not on one fixed provider or model identity.
Collaborative
Let specialist models share context, critique, and verify each other's work.
Controlled
Give capabilities distinct permissions and resource boundaries, including sandboxing.
Edge-ready
Explore useful agent runtimes for laptops, phones, Raspberry Pi, and constrained devices.
05 / First experiment
Measure the idea before scaling it.
The first prototype is designed to test where dynamic capability management helps, and where it does not.
- 01
Build a minimal Python runtime API.
- 02
Connect multiple interchangeable models.
- 03
Select capabilities through an orchestrator.
- 04
Implement active, sleeping, and unloaded states.
- 05
Share task context across capabilities.
- 06
Measure latency, memory, cost, and model calls.
- 07
Benchmark against single-model and fixed multi-model agents.
The central research question
When should an intelligence activate, stay loaded, sleep, unload, or be replaced?
Morph is not only about model routing. It studies the dynamic lifecycle of a population of models inside an agent, and compares the results against single-model and fixed multi-model baselines.
Back to Phylogeny projects