Runtime, Harness & Cloud Model Accountability Ratings
Factual 5-axis A1-style accountability scorecard across 20 AI harnesses, cloud model endpoints (including Ollama Cloud muse-spark-open & glm-5.3-flash at https://ollama.com/v1), and execution substrates. Machine-readable feed: /ratings.json.
| # | Runtime / Harness / Cloud Endpoint | Grade | Cost Attr | Spend Caps | Audit Log | CAS Safety | Portability | Factual Verification Notes |
|---|---|---|---|---|---|---|---|---|
| #1 | Teleon THP v1.0 + D1 Fabric Accountability & Queue Substrate · https://api.teleon.dev (THP v1.0) |
A1 (98/100) | 20/20 | 20/20 | 20/20 | 20/20 | 18/20 | Native X-Teleon-Surface-Id attribution, monotonic lease_epoch queue fencing, expected_version CAS stores, and SWHID receipts. |
| #2 | Cloudflare Workers + D1 SQLite Edge Execution Runtime · V8 Isolates + D1 Transactional SQLite |
A1 (94/100) | 19/20 | 19/20 | 18/20 | 20/20 | 18/20 | Per-worker CPU/wall limits, zero-ambient filesystem isolation, parameterized SQLite CAS writes, and deterministic WebCrypto. |
| #3 | Ollama Cloud (muse-spark-open) Cloud Model Endpoint · https://ollama.com/v1 · ollama-cloud/muse-spark-open |
A1 (93/100) | 19/20 | 18/20 | 19/20 | 18/20 | 19/20 | OpenAI-compatible cloud API with exact prompt/completion token telemetry, high-reasoning drift calibration, and portable model IDs. |
| #4 | Ollama Cloud (glm-5.3-flash) Cloud Model Endpoint · https://ollama.com/v1 · ollama-cloud/glm-5.3-flash |
A1 (92/100) | 19/20 | 19/20 | 18/20 | 17/20 | 19/20 | Sub-second cloud flash throughput over https://ollama.com/v1 with deterministic JSON tool-call schemas and low per-task spend. |
| #5 | OpenCode (Ollama Cloud + MCP Mode) Agent Coding & Fleet Harness · OpenCode CLI + https://ollama.com/v1 + Streamable HTTP MCP |
A1 (91/100) | 18/20 | 18/20 | 19/20 | 17/20 | 19/20 | First-class Streamable HTTP MCP client, clean provider routing to Ollama Cloud models, and explicit per-session tool permission gates. |
| #6 | Google Antigravity (AGY 2.0) Agentic Engineering Harness · Antigravity Workspace + Structured JSONL Transcripts |
A1 (91/100) | 18/20 | 18/20 | 20/20 | 17/20 | 18/20 | Full step-indexed JSONL execution transcripts, subagent conversation isolation, MCP native bindings, and explicit rule enforcement. |
| #7 | Anthropic Claude Code Agent Coding Harness · Claude Code CLI + .mcp.json + CLAUDE.md |
A2 (86/100) | 18/20 | 17/20 | 17/20 | 16/20 | 18/20 | Granular tool allowlists, session cost counters, native MCP transport, and deterministic hook interception before shell writes. |
| #8 | OpenAI Codex CLI Sandboxed Coding Harness · Codex CLI + Kernel Seatbelt/Landlock Sandbox |
A2 (85/100) | 17/20 | 18/20 | 17/20 | 16/20 | 17/20 | OS-enforced network and write sandboxing, AGENTS.md instruction hierarchy, and TOML-configured MCP server integration. |
| #9 | Temporal Cloud Worker Fleet Durable Execution Runtime · Temporal Event History + Deterministic Replay |
A2 (86/100) | 17/20 | 17/20 | 19/20 | 18/20 | 15/20 | Immutable workflow event histories, task-queue rate limits, and replay determinism; pairs cleanly with SWHID artifact commits. |
| #10 | Anthropic Messages API (Claude 4.6) Cloud Model Endpoint · https://api.anthropic.com/v1/messages |
A2 (84/100) | 18/20 | 17/20 | 17/20 | 15/20 | 17/20 | Workspace-scoped API keys with monthly USD spend caps, cache-token accounting, and strict tool-use schema validation. |
| #11 | Google Vertex AI (Gemini 2.5 Pro) Cloud Model Endpoint · Vertex AI IAM + Cloud Audit Logs |
A2 (84/100) | 18/20 | 17/20 | 18/20 | 15/20 | 16/20 | IAM service-account attribution, Cloud Audit Logs data-access receipts, and quota caps per project and region. |
| #12 | OpenAI Responses API (GPT-5) Cloud Model Endpoint · https://api.openai.com/v1/responses |
A2 (83/100) | 18/20 | 17/20 | 17/20 | 14/20 | 17/20 | Project-level budget limits, native remote MCP tool calls, and structured response IDs; requires external CAS store for state. |
| #13 | Fly.io Machines + LiteFS MicroVM Container Runtime · Firecracker MicroVMs + Regional Volumes |
B1 (79/100) | 16/20 | 16/20 | 15/20 | 16/20 | 16/20 | Hardware-virtualized microVM boundaries and stop-on-idle leases; requires operator guardrails against volume deletion. |
| #14 | GitHub Copilot Coding Agent Repository Workflow Harness · GitHub Actions Ephemeral Runner + PR Gate |
B1 (78/100) | 16/20 | 16/20 | 17/20 | 15/20 | 14/20 | Runs inside ephemeral Actions containers with git commit provenance and mandatory human pull-request review gates. |
| #15 | Cursor Background Agents Cloud IDE Harness · Remote VM Branch Worker + .cursor/mcp.json |
B1 (77/100) | 16/20 | 15/20 | 16/20 | 15/20 | 15/20 | Per-agent spend controls and isolated git branch checkout; outbound non-git side effects still require MCP fencing. |
| #16 | AWS Bedrock Agents Enterprise Cloud Runtime · AWS IAM + CloudTrail + Action Groups |
B1 (76/100) | 17/20 | 16/20 | 17/20 | 14/20 | 12/20 | Strong CloudTrail auditability and IAM role scoping, with lower cross-cloud identity portability outside AWS. |
| #17 | Vercel AI SDK + Fluid Compute Serverless Agent Runtime · TypeScript Streaming Tool Loop |
B2 (72/100) | 15/20 | 14/20 | 14/20 | 13/20 | 16/20 | Clean multi-provider model abstraction and step count caps (maxSteps); relies on external DB for lease and CAS safety. |
| #18 | LangGraph Cloud Platform Graph Orchestration Runtime · Checkpointed State Graph + Interrupts |
B2 (71/100) | 14/20 | 14/20 | 15/20 | 15/20 | 13/20 | Supports human-in-the-loop graph interrupts and thread checkpoints; identity portability remains coupled to framework state. |
| #19 | CrewAI AMP Cloud Multi-Agent Role Framework · Role-Delegated Python Agent Crew |
C1 (61/100) | 12/20 | 12/20 | 13/20 | 11/20 | 13/20 | Rapid multi-agent prototyping, but peer delegation requires external THP clamping to prevent Chatter-loop token burn. |
| #20 | Unfenced Raw Cron + Shell Webhooks Ad-Hoc Baseline (Control Group) · Ambient ENV Secrets + Unversioned SQL Writes |
D1 (28/100) | 5/20 | 4/20 | 6/20 | 5/20 | 8/20 | Zero per-agent cost attribution, shared ambient API tokens, split-brain duplicate runs on retry, and zero SWHID provenance. |