Multi-agent orchestration
Route each task to the model topology that fits it — sequential, fan-out, debate, or hierarchy. Weighted routing, retries, and failover run in the plane, not in your agent code.
POST /v1/orchestration/route
Agent control plane · v2.4
Run autonomous agents like you run production.
aetherc is the control plane for agent fleets. Route work across model topologies, trace every reasoning step and tool call, hold agents inside budget and permission boundaries, and take the controls mid-run when a decision needs a human.
Execution topology
6 nodes · 7 edges live
Routing latency p95−12%
63ms
Throughputtok/s
18.4k
Spend by hour72% of cap
| run | agent | state | steps | latency | spend |
|---|---|---|---|---|---|
| r-8f31·02 | Data Extractor | running | 12/18 | 41ms | $0.84 |
| r-8f2e·07 | Code Refactorer | running | 41/60 | 88ms | $3.20 |
| r-8f2a·01 | Safety Evaluator | intervention | 6/9 | — | $0.11 |
| r-8f27·04 | Vendor Parser | throttled | 3/22 | 210ms | $1.92 |
Core surfaces
Four controls, one plane.
Each surface is a first-class API and a first-class view. What you can see in the console, you can call from the SDK — same primitives, same guarantees.
Real-time observability
Every reasoning step, tool call, token, and retry lands in a trace you can replay. Filter a run down to the one call that blew the latency budget, then diff it against the last green run.
GET /v1/telemetry/tracePolicy & guardrails
Declare spend caps, tool permissions, data-residency rules, and egress allowlists once. The plane enforces them at call time and halts the run before a boundary is crossed.
PUT /v1/policy/guardHuman in the loop
Pause a run mid-step, inspect the pending action, edit the arguments, then approve or deny. Reviews route to the right person by policy and expire into a safe default.
POST /v1/human/interrupt
Interactive · live state
Take the controls.
A working slice of the console, running in this page. Filter the fleet, hold an agent mid-step, and resolve the review the Safety Evaluator is about to raise.
3active agents
0tokens / min
$0.00spend this run
0interventions
Execution log
all agents
Simulated fleet · no data leaves this page
Docs: controls/interrupt
Deployment
Your models, your network.
The control plane is stateless by design. Run it managed, split it across your VPC, or take the whole thing behind an air gap — the agent contract does not change.
| Mode | Control plane | Data plane | Model runners | Egress | Attestation |
|---|---|---|---|---|---|
| Managed cloudaetherc.com | aetherc | aetherc | Hosted frontier + OSS | Allowlist | SOC 2 Type II |
| Hybridbring your own compute | aetherc | your VPC | Any endpoint you expose | Private link | SOC 2 + BYOK |
| Self-hostedhelm · terraform | your cluster | your cluster | Local + remote | Your policy | your controls |
| Air-gappedoffline bundle | your cluster | your cluster | Local runners only | none | signed images |
full support
partial / your responsibility
not applicable
Model runners
Auto-discovered on the node, health-checked every 10s.
OllamaDocker Model RunnerNVIDIA NIM
vLLMllama.cppBedrockVertex AI
Agent frameworks
Drop-in adapters. Keep your graph, gain the plane.
Claude Agent SDKLangGraphCrewAI
AutoGenOpenAI Agents SDKTemporalRay
Telemetry sinks
OTLP native. Traces stream while the run is still open.
OpenTelemetryDatadogGrafana
HoneycombKafkaS3Postgres
Start now
Put a fleet on the board.
The demo workspace ships with three live agents, a seeded policy set, and a trace history you can break. No install, no card.