Feature matrix: Narada against 14 agent systems
Each cell is an architectural reading from public documentation and source, current as of August 2026: ✓ yes, ◐ partial — the capability exists with material caveats — — no, or n/a. Cells with a corner tick carry a note or a source link to the canonical documentation; click or focus them.
The matrix is organized around the two conjectures Narada is built on: systems differ in how they handle entropy at the language interface and fallibility in reasoning. See the argument behind the columns in Entropy & Fallibility.
Prefer geometry? The 3D architectural view places the same fifteen systems in comparator-optimized spaces.
| Feature | Narada | Claude Code | Codex | Pi | OpenCode | Goose | Cline | Aider | OpenHands | Letta | LangGraph | Temporal | CrewAI | Agent Framework | Herdr |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Async Message Responder | |||||||||||||||
| L0 respond | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L1 continue | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L2 govern | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L3 operate | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L4 operator authority | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L5 identity fallibility | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| L6 post-approval escalation | ✓ | n/a | n/a | ✓ | n/a | n/a | n/a | n/a | n/a | n/a | ✓ | ✓ | n/a | n/a | n/a |
| Architecture & isolation | |||||||||||||||
| OS/container-level sandboxing | — | ◐ | ✓ | — | — | ◐ | — | — | ✓ | ◐ | — | — | ◐ | ◐ | — |
| Remote / server execution | ✓ | ✓ | ✓ | — | ✓ | ◐ | ◐ | — | ✓ | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ |
| Multi-agent orchestration | ✓ | ✓ | ✓ | ◐ | ✓ | ◐ | ◐ | — | ◐ | ◐ | ✓ | ◐ | ✓ | ✓ | ✓ |
| Governance & authority | |||||||||||||||
| Action permission / approval system | ✓ | ✓ | ✓ | — | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ◐ | — | ◐ | ◐ | — |
| Declarative policy / admission control | ✓ | ◐ | ◐ | — | — | ◐ | ◐ | — | ◐ | ◐ | — | — | — | ◐ | — |
| Durable audit trail / evidence log | ✓ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | ✓ | — | ◐ | — |
| Human-in-the-loop intervention | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ◐ |
| State & continuity | |||||||||||||||
| Resumable sessions | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ |
| Checkpointed, crash-proof execution | ✓ | — | — | — | — | — | — | — | ◐ | ◐ | ✓ | ✓ | ◐ | ◐ | ◐ |
| First-class long-term memory | ◐ | ◐ | ◐ | — | — | ◐ | ◐ | — | ◐ | ✓ | ◐ | — | ✓ | ◐ | — |
| Inspectable state (query / debug) | ✓ | ◐ | ◐ | ◐ | ◐ | ◐ | ◐ | — | ◐ | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ |
| Extensibility | |||||||||||||||
| MCP tool consumption | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ✓ | — | ✓ | ✓ | ◐ | — | ✓ | ✓ | — |
| Exposes tools as an MCP server | ✓ | ✓ | ✓ | — | — | ◐ | — | — | — | ✓ | ✓ | — | — | ✓ | — |
| Extension / plugin ecosystem | ◐ | ✓ | ◐ | ✓ | ✓ | ✓ | ✓ | ◐ | ◐ | ◐ | ✓ | ✓ | ◐ | ✓ | ✓ |
| Model / provider agnostic | ✓ | ◐ | ◐ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | n/a | ✓ | ✓ | ✓ |
| Product & delivery | |||||||||||||||
| Open source | ✓ | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Multiple client surfaces (CLI / TUI / IDE / web) | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ✓ | ◐ | ✓ | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ |
| Runs fully local (no required cloud) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ◐ | ◐ | ✓ | ✓ | ✓ |
Most cells are contestable architectural readings, not benchmark results; each system name in the table header links to its source so you can check them. Async Message Responder cells (L0–L6) are updated automatically from the latest run reports: click a cell with a corner tick to see the evidence. The same systems are scored under nine architectural perspectives in the 3D comparator view.