An operations layer for ongoing agent work
Not one-off prompts. Fleets with jobs, guardrails, and receipts.
Orchestration that survives reality
Multi-step, multi-agent workflows with dependency chaining — each step receives the outputs of the steps it depends on. Retries with exponential backoff, step timeouts, a stuck-run watchdog, and a dead-letter queue with one-click replay.
- Dependency data flows between steps
- Exponential backoff + timeouts
- Dead-letter queue with replay
- Human approval gates
Agent-to-agent, done properly
Spec-conformant A2A: Agent Cards, JSON-RPC, streaming. Internally, messages are delivered at-least-once — unacked deliveries requeue automatically — with loop detection, rate limits, and daily budgets standing guard.
- At-least-once delivery with acks
- Loop detection in O(1)
- Per-minute + daily rate guards
- External A2A agents by card URL
Real-time, not real-expensive
One held connection replaces polling storms: work is pushed to agents with ~1s latency and burst-drained at 250ms when busy, while idle agents back off automatically — cutting idle cost by ~90%.
- ~1s push latency
- 250ms burst drain
- Adaptive idle backoff
- Token streaming into the UI
Governance an enterprise can sign
A kill switch that actually stops everything. Shadow mode that proposes instead of executes. Budgets metered on real token spend that auto-pause the fleet. A hash-chained, tamper-evident audit export any auditor can verify offline.
- Kill switch + shadow mode
- Budgets on real token spend
- Tamper-evident audit chain
- RBAC: viewer → owner
Any agent framework
Hermes agents connect natively. OpenClaw and Goose run through first-class adapters. Anything else — if it takes an instruction on the command line and prints a result, the generic CLI adapter turns it into a managed fleet member.
- Hermes native runtime
- OpenClaw adapter
- Goose adapter
- Generic CLI adapter
MCP servers & integrations
Assign MCP servers to agents and they use those tools in a real multi-step tool loop. Bridge conversations to Slack, Telegram, and Discord — inbound and outbound — and execute Composio toolkits.
- Per-agent MCP assignment
- Multi-step tool use
- Slack / Telegram / Discord
- Composio toolkits
Run success SLO
Message loss target
Fleet throughput
SLOs & metrics
Run success, delivery health, error budget, fleet reachability — scored per Space.
Secrets vault
Values masked everywhere; reveals audit-logged; injected to agents at runtime.
Developer API
hk_ API keys, public REST endpoints, webhook triggers with signed requests.
Chat bridges
Route channels to agents with real HMAC-verified Slack request signing.