Narada is not a single score on a universal benchmark. Its shape depends on the question being asked of an agent system.
The visualization places six systems in a three-dimensional space, then changes the meaning of the three axes when a different system becomes the comparator. Each comparator defines the virtues it optimizes for. Narada is shown alongside Pi, OpenCode, OpenHands, Letta, and LangGraph so that its relative strengths and tradeoffs remain visible rather than collapsing into one ranking.
The systems
- Pi — a minimal, deeply extensible local coding-agent runtime whose session, tools, extensions, and primary interaction remain close together.
- OpenCode — a coding-agent backend behind a server boundary, designed to support multiple clients while retaining a coherent product runtime.
- OpenHands — a production-oriented agent platform emphasizing controlled remote execution, isolated workspaces, and operational delivery.
- Letta — a persistent-agent architecture centered on durable identity, memory continuity, and inspectable server-owned state.
- LangGraph — a durable execution system centered on explicit graphs, queued runs, checkpoints, workers, and composable orchestration.
- Narada — a proposed governed agent runtime separating durable session control, provider execution, capability admission, evidence, and replaceable clients.
Six comparators
| Comparator | X axis | Y axis | Z axis |
|---|---|---|---|
| Pi | Architectural compactness | End-user plasticity | Interaction immediacy |
| OpenCode | Client/server clarity | Multi-surface readiness | Coding-product coherence |
| OpenHands | Workspace isolation | Remote execution operability | Production delivery |
| Letta | Identity persistence | Memory continuity | State inspectability |
| LangGraph | Workflow explicitness | Durable execution | Orchestration composability |
| Narada | Continuity externalization | Execution disaggregation | Authority formalization |
Under Pi’s perspective, Pi sits at the (10,10,10) corner and Narada scores lower because Narada deliberately spreads agency across separate components. Under Narada’s perspective, the ranking inverts: Narada occupies the optimum, while Pi looks compact but insufficiently governed.
How to read it
Select a perspective to make its comparator the canonical point at (10,10,10). The other points show how each system is judged relative to the comparator’s chosen virtues. Select a point to edit its coordinates locally, then use the reset controls to return to the canonical values.
The coordinates are ordinal architectural judgments on a 0 to 10 scale. They are a thinking instrument, not empirical measurements.
Why this framing matters
A single ranking of agent systems would be misleading because the systems optimize for different architectural concerns. Pi’s compactness is a virtue for local coding but a liability for operational governance. LangGraph’s explicit workflow structure is powerful for orchestration but may feel heavy for a user who wants immediate, plastic interaction.
The graph is intentionally comparative. It does not claim that Narada is universally better than the other systems. It shows that Narada’s distinctive claim is strongest where continuity, execution boundaries, and authority are treated as separate architectural concerns. Under another comparator’s priorities, a different system can occupy the optimum.
Residual
The coordinates need future grounding in implementation evidence, operational traces, and explicit evaluation criteria. Until then, the graph is a disciplined hypothesis: useful because it exposes which virtues are being optimized, provisional because the judgments remain contestable.