Contain every AI agent. Lose no speed.
Orion runs your agents inside a 14-layer containment and audit pipeline at no measurable speed cost, and its accelerators help those agents produce more working code on hard tasks.
Measured on a single consumer PC. Every number on this page names its test.
Many agents at once is easy. Trusting them is the hard part.
Teams that run agents on customer data, money or regulated work run into the same four failures. Most either accept the risk or add checks that slow everything down, and then switch the checks off.
Runaway loops
An agent repeats itself forever, burning compute and budget.
Answered by MaxTurnGovernor, PopulationController
Data leaks between agents
One agent sees another’s data or customer records.
Answered by AgentContainer, DataSwitch, SandboxEnforcer
Hijacking by bad prompts
Injected text makes an agent ignore its instructions.
Answered by SafetyWatcher, EvaluatorGate, ControlEngine Bridge
Work nobody can verify
Afterwards, no one can prove what an agent did or why.
Answered by MultiChain, a SHA-256 tamper-evident audit log
Safety that costs nothing, measured.
The same model on the same machine, run with and without Orion. Raw means the model alone.
Same generation speed
Average tokens per second with all 14 containment layers and all 5 accelerators on, versus the bare model, across 7 load levels.
Requests dropped, 10 to 100 agents
Zero lost requests up to 100 concurrent agents on one RTX 4060 PC. Under overload (200 to 1,000 agents), Orion lost 0.4 to 1.6 points fewer requests than the bare model at every level.
Higher code scores on hard tasks
With accelerators on, the average score across 10 hard programming tasks rose from 80.9 to 86.1. The biggest gains came on the hardest work: Garbage Collector +27, Lock-Free Queue +11, HTTP Server +10.
Complete, runnable solutions
Orion produced 7 complete solutions that run, against 5 for the bare model, from the same 10 tasks.
Lost requests by load
| Agents | 10 | 20 | 50 | 100 | 200 | 500 | 1,000 |
|---|---|---|---|---|---|---|---|
| Raw | 0% | 0% | 0% | 0% | 20.2% | 55.1% | 73.3% |
| Orion | 0% | 0% | 0% | 0% | 19.1% | 54.7% | 71.7% |
Above 100 agents, the limit is the graphics card, not Orion: the model server handles 8 requests at once in both modes. Orion scales as far as your hardware does and never becomes the bottleneck.
What we don’t claim
- Orion is not faster than the bare model. It is the same speed.
- Orion doesn’t fix overload. Past 100 agents on one PC, both setups lose most requests.
- The containment layers don’t make code better. The accelerators do.
- Containment has not yet been tested against live attacks. A red-team benchmark is built and scheduled.
Test setup (Benchmark Report v3, October 2026): NVIDIA RTX 4060 8 GB, AMD Ryzen 9 9900X, 32 GB DDR5, Windows 10, llama3.2:3b via Ollama with 8 parallel slots. Load levels of 10 to 100 agents ran for 5 minutes each; 200 to 1,000 for 10 minutes each. Each level and task ran once; repeat runs with ranges are planned.
What happens to one request
Every request from an agent passes through Orion on its way to the model and again on its way back. The whole trip adds about 6 milliseconds.
Admit. ControlEngine Bridge receives the request. PopulationController and MaxTurnGovernor check the agent is allowed to act.
Isolate. SandboxEnforcer, ContainerManager, AgentContainer and DataSwitch make sure the request only touches that agent’s own data.
Brief. Watcher attaches relevant lessons from the Whiteboard, and Navigator attaches the build plan.
Draft. Shuttle tries a fast draft on a small 1B model. If it fails a quick check, the main model takes over with the draft and the reasons it failed.
Check. SafetyWatcher and EvaluatorGate review the answer before it returns.
Record. MultiChain logs every step. If anything goes wrong, Prison quarantines the agent, BurnRegistry records a strike, and DoubleFailsafeWatcher can lock the swarm down.
Learn. When the agent finishes or crashes, Last Gasp writes its lesson back to the Whiteboard for the next agent.
14 containment components
- DoubleFailsafeWatcher
- Two independent containment layers. If either fails, all agents lock down.
- MultiChain
- Immutable SHA-256 audit log of every agent action.
- Prison
- Quarantines a misbehaving agent without stopping the swarm.
- BurnRegistry
- Tracks resource use and strikes; removes repeat offenders.
- RewardRegime
- Rewards agents for completing tasks correctly.
- PopulationController
- Manages spawning and termination.
- DataSwitch
- Routes data between agents and blocks unauthorised cross-talk.
- SafetyWatcher
- Monitors outputs in real time for policy violations.
- MaxTurnGovernor
- Limits turns per agent to stop infinite loops.
- SandboxEnforcer
- File and network isolation for each agent.
- EvaluatorGate
- Lets only properly formed responses through.
- ContainerManager
- An isolated work environment per agent.
- AgentContainer
- Keeps agents from seeing each other’s data.
- ControlEngine Bridge
- Connects the model to every component, so nothing bypasses the guardrails.
5 accelerators
- Last Gasp Protocol
- Every agent, finished or crashed, leaves a note with partial work and a lesson.
- Whiteboard
- A shared board where agents learn from earlier agents’ notes.
- Watcher
- Passes each agent only the notes relevant to its task.
- Navigator
- A small planner model on the CPU that keeps a build plan per task.
- Shuttle
- A fast first draft from a small model, checked before the main model steps in.
Working today, and what’s next
Live now
- All 14 containment components
- All 5 accelerators
- SHA-256 tamper-evident audit trail, anchored on Arbitrum testnet
- PBKDF2 / Fernet encrypted vault
- Per-agent container isolation (Docker)
Planned
- Live red-team results against injection, leaks and runaway loops
- Head-to-head runs against leading agent frameworks
- Zero-knowledge proofs and XMSS post-quantum signatures
- Audit anchoring on mainnet
- VM-level isolation per agent
Run your agents inside Orion
We’re opening a small number of design-partner pilots for teams running agents on sensitive or regulated work. Tell us what your agents do and we’ll be in touch.