Architecture
A swarm needs contracts, not just personas
Agent names and prompts do not define responsibility. Durable swarms need inputs, outputs, authority, state, completion, failure, and evidence contracts.
Insights
Operating guidance for durable machine intelligence
The core question is not whether an agent appears intelligent. It is whether the surrounding system remains understandable, governable, recoverable, and useful as it changes.
EnterpriseGovernmentPartners
Research agenda
Current themes
Each theme is framed as an operating question rather than a technology prediction.
Control
The control plane is where governance becomes real
Policy must be evaluated at identity, delegation, tool, data, approval, and action boundaries—not only documented.
Evaluation
Evaluate the coordination path
A correct final result can hide redundant work, unsafe tools, runaway retries, policy bypass, or irreproducible state.
Memory
Memory is an ownership problem
Shared context creates value and risk. Teams must distinguish transient state, durable memory, knowledge, evidence, and records.
Authority
Human-in-the-loop is not an authority model
A human presence is insufficient unless review criteria, authority, evidence, timing, and escalation are explicit.
Design
More agents do not guarantee more intelligence
Additional roles increase coordination cost and failure surfaces. Use the smallest topology that improves the decision.
Operations
Replay is a reliability primitive
A system that cannot reconstruct its task graph, state, versions, calls, policies, and approvals cannot learn safely from failure.
Portability
Provider portability starts above the model API
Portable systems express behavior in contracts, controls, fixtures, evidence, and routing—not in framework-specific glue.
Risk
Autonomy should expand by action class
Grant authority gradually based on consequence, evidence, reversibility, detection, and recovery—not with one global autonomy setting.
Start with a bounded decision
Turn an insight into a bounded architecture or evaluation decision.
Use the service ladder when the system needs evidence, not another conceptual discussion.