State fragments
Goals, decisions and evidence scatter across prompts, logs and tools.
Agent continuity and execution infrastructure
The model can change.
The thread does not.
Gimbyl is being built to keep an agent’s canonical state—its goals, permissions, evidence, memory and task continuity—independent of any single model or provider.
A model-portable agent runtime and governed context compiler from CuraIQ Labs.
MODEL / 01
Reason
Context compiled
MODEL / 02
Act
Policy boundary
MODEL / 03
Resume
State portable
CANONICALTHREAD
CONTINUITY / REFERENCE ARCHITECTURE
01 Model-portable
02 Governed context
03 Context compiled
04 Inspectable
01 / THE PROBLEM
Continuity, not another model
Models are replaceable.
Work should not be.
Agents lose the plot when context is trapped inside a model session, a process or a device. Gimbyl is designed to separate the durable thread from the intelligence temporarily acting on it.
Models move
A more suitable model for the next task may not be the one that began the work.
Control fades
Without an independent record, permissions and provenance become difficult to inspect.
02 / THE RUNTIME
One durable thread
A stable layer beneath a changing intelligence stack.
Gimbyl is being built to maintain canonical state, compile a purpose-built, authorised context for each invocation, and route work to a suitable model without tying continuity to one provider.
INTELLIGENCE
Frontier model
Specialist model
Local model
SWAPPABLE
CONTINUITY
GIMBYL
DURABLE
EXECUTION
Tools
Data
Devices
Workflows
DISTRIBUTED
Canonical state
A provider-neutral record of goals, permissions, evidence, memory and task continuity—independent of the current model session.
Governed context
A purpose-built, authorised context is compiled from canonical state for each model invocation.
Portable execution
Route work across suitable cloud-hosted and local models while canonical work state remains independent of the provider.
03 / THE LOOP
Continuity in motion
Every invocation advances the same thread.
Gimbyl is being built to turn fragmented agent activity into one inspectable sequence. The aim is to change the model, move the process or resume on another device without reconstructing the work by hand.
- 01ObserveCapture outcomes, evidence and state transitions.
- 02CompileAssemble the minimum authorised context for the task.
- 03RouteSelect the model and environment suited to the work.
- 04ResumeCommit the result back to the durable thread.
04 / PRINCIPLES
Built for agency that lasts
Continuity without lock-in.
Control without friction.
No model owns the agent.
Intelligence can evolve without forcing the work to restart.
Authority remains explicit.
Permissions and purpose boundaries can be inspected as tasks move across systems.
Evidence stays attached.
Decisions and state changes remain available for review.
Local when needed. Cloud when useful.
Execution can shift without severing the thread.
Private preview
Keep the thread.
We are inviting a small group of teams working on long-running, high-trust agents to help test the Gimbyl thesis.