Comparison tables for OpenSentience runtime governance relative to agent frameworks, guardrail libraries, and harness architectures.
Yes/No indicates presence or absence of the feature as a first-class, enforceable mechanism (not just documentation or convention).
None means the system has no equivalent concept.
Partial or qualified entries describe limited coverage.
Comparisons reflect each system's public documentation as of early 2026. Features may change.
OpenSentience protocols (OS-001 through OS-010) are defined in opensentience.org/docs/spec/README.md. OS-009 (PRISM) and OS-010 (PULSE) have separate specification files.
| Feature | OpenSentience | CrewAI | LangGraph | AutoGen | Guardrails AI | NeMo Guardrails | AWS Bedrock |
|---|---|---|---|---|---|---|---|
| Permission model | Deny-by-default, ETS-cached, 4 categories (filesystem, network, tool, graph) | None | None | None | Input/output validation only | Input/output rails | IAM-based (cloud) |
| Lifecycle management | 5 states (installed, enabled, running, disabled, removed) | None | None | None | None | None | Cloud-managed |
| Graduated autonomy | 3 levels (observe / advise / act) | None | None | None | None | None | None |
| Audit trail | Append-only, immutable, batched, pluggable backends | None | None | None | Per-call logs | Per-call logs | CloudTrail |
| Pipeline enforcement | OS-008: prerequisite constraints, retrieve-before-act | None | Static graph routing | Conversation flow | None | Topical flow | None |
| Quality gates | OS-008: separate evaluator context, adversarial grading | None | None | None | Validators | None | None |
| Sprint contracts | OS-008: planner, generator, evaluator with acceptance criteria | Task assignment | None | None | None | None | None |
| Context management | OS-008: 60% threshold, compaction, Graphonomous overflow | None | Checkpoints | None | None | None | None |
| Cognitive grounding | 10 protocols (8 cognitive primitives + PRISM diagnostic + PULSE temporal), each mapped to cognitive science | None | None | None | None | None | None |
| Loop manifest standard | OS-010 PULSE: declarative manifest with 5 phase kinds, 5 cross-loop tokens, 12-test conformance suite | None | None | None | None | None | None |
| Continual-learning benchmark | OS-009 PRISM: 9 CL dimensions, BYOR ingestion, IRT calibration, leaderboards | None | None | None | None | None | None |
| Runtime overhead | < 1% CPU, < 2us permission check | N/A | N/A | N/A | Per-call latency | Per-call latency | Cloud latency |
| Architecture | Elixir/OTP hex package (shim) | Python framework | Python framework | Python/.NET framework | Python library | Python library | Cloud service |
| Portability | Apache 2.0, any OTP child_spec | Framework-locked | Framework-locked | Framework-locked | Provider-agnostic | Provider-agnostic | AWS-locked |
| Multi-agent governance | Per-agent permissions + autonomy + Delegatic policy | Role-based | Graph-based | Conversation-based | N/A | N/A | IAM-based |
| Protocol | Layer | What it governs | Enforcement type |
|---|---|---|---|
| OS-001 | Memory | How knowledge is stored, consolidated, retrieved | Spec (Graphonomous enforces) |
| OS-002 | Routing | When to fast-path vs. deliberate | Spec (Graphonomous enforces) |
| OS-003 | Deliberation | How agents reach consensus | Spec (AgenTroMatic enforces) |
| OS-004 | Attention | What gets priority | Spec (Graphonomous enforces) |
| OS-005 | Resources | Which model tier handles a task | Spec (Graphonomous/Agentelic enforce) |
| OS-006 | Governance | Permissions, lifecycle, autonomy | Runtime (open_sentience hex) |
| OS-007 | Security | Adversarial threats | Runtime (planned) |
| OS-008 | Harness | Pipeline ordering, quality gates, sprint contracts | Runtime (planned) |
| OS-009 | Diagnostic | How well a closed memory loop actually learns over time | Runtime (PRISM Elixir/OTP, Fly.io, 6 MCP machines) |
| OS-010 | Temporal | How loops cycle, nest, and signal across boundaries | Manifest standard (no required runtime; JSON Schema validation + 12-test conformance suite) |
| Feature | Anthropic Harness | OpenAI Codex Harness | LangChain Deep Agents | OS-008 |
|---|---|---|---|---|
| Generator-evaluator separation | Yes (3-agent) | Implied (CI gate) | No | Yes (capability contracts) |
| Sprint decomposition | Ad-hoc | Implicit | No | Formal (SprintController + acceptance criteria) |
| Context management | Filesystem offload | Filesystem | Agent delegation | Graphonomous knowledge graph + compaction |
| Prerequisite enforcement | Prompt-based | CI invariants | None | Runtime (PipelineEnforcer) |
| Governance integration | None | None | None | [&] Protocol contracts + Delegatic policy |
| Topology-aware routing | None | None | None | kappa-routing (OS-002) |
| Coverage-driven dispatch | None | None | None | Coverage.recommend: act / learn / escalate |
| Model tier adaptation | Fixed model | Fixed model | Fixed model | 3-tier with graceful degradation |