Hannah Composer
The super-orchestrator that turns plain-language intent into an executable agent plan
Hannah Composer is the front door of the Hannah Imagineer ecosystem. It decomposes a single user prompt into a typed plan, dispatches the work across specialist subagents (Hermes for ops, OpenClaw for chat-native execution, Shared Mind for retrieval), and supervises the run end-to-end with checkpointing, retries, and human approval gates.
Overview
Composer is the planner, not just a router. Given an intent, it produces a structured plan — DAG of subagent calls, expected artifacts, approval gates — and persists it before execution. Every step is observable, every artifact is addressable, and any subagent can be swapped without rewriting the plan. The Composer surface is what powers the marketing site, the in-app onboarding flow, and the customer-facing Imagineer assistant.
Primitives
Intent decomposition into typed, addressable subagent plans
Multi-agent dispatch with checkpointed state
Human-in-the-loop approval gates on destructive steps
Live plan editing — pause, branch, retry from any node
Cross-agent memory hand-off via Shared Mind
Built-in telemetry and run-replay for every plan
Outcomes
- 01One conversational surface replaces a dozen bespoke automations
- 02Plans are inspectable artifacts — no black-box agent runs
- 03Customers can compose new workflows without writing code
- 04Every subagent improvement compounds across every plan
Integrations
Autonomy & guardrails
- Cross-tenant subagent dispatch
- Spending-impacting actions above the per-plan budget cap
- External communications outside the configured perimeter
Guardrails & requirements
Guardrails
- Plans are immutable once dispatched; edits create a new plan version
- Destructive subagent calls require explicit approval
- Subagent egress is filtered through the Privacy Router
- Full run trace persisted for audit and replay
Requirements
- Hannah Imagineer workspace (managed) OR self-hosted Hannah stack
- Outbound HTTPS to chosen frontier-model provider
Technical specifications
Runtime
- Harness
- Hannah Composer Runtime
- Deployment
- Managed (Hannah Imagineer cloud) · Self-hosted (single tenant)
- Data residency
- Customer-controlled — Composer stores plans and scratchpads in the customer Supabase instance
- License
- Commercial · Hannah Imagineer EULA
- Version
- v1.0
Models & tooling
- Models
- claude-opus-4-7claude-sonnet-4-6gpt-5
- Tooling
- plan-dslsubagent-dispatchercheckpoint-storeapproval-gate-enginerun-replay
Reliability & limits
- Latency SLA
- ~1.2s p50 plan compile · per-step latency dominated by subagent
- Rate limits
- 60 plans/min per workspace
Security & compliance
- Auth model
- Supabase auth · SSO via Okta / Google Workspace
- Compliance
- SOC 2 Type II (in process)GDPR
Evals & benchmarks
Composer-native benchmarks (plan-replay determinism, approval-acceptance rate) ship in the next release. Until then we publish the upstream harness numbers from the OpenHands SDK paper as the credible ceiling: any Composer plan that bottoms out in code editing inherits the harness performance, and we report Composer's overhead delta on top.
| Benchmark | Model | Score | Notes | Source |
|---|---|---|---|---|
SWE-Bench Verified | Claude Sonnet 4.5 | 72.8% | Upstream harness baseline reported by OpenHands SDK paper. Composer adds orchestration overhead; full agent-overhead delta benchmark is in progress. | link |
GAIA (validation) | Claude Sonnet 4.5 | 67.9% | Upstream harness baseline reported by OpenHands SDK paper. Cited as the closest credible proxy for Composer's tool-use ceiling. | link |
Architecture notes
Memory architecture
Composer maintains a per-tenant plan store and a per-run scratchpad. Long-term context lives in Shared Mind (the second-brain RAG layer); short-term context lives in the run scratchpad and is garbage-collected once a plan terminates.
Context strategy
On dispatch, Composer compiles the plan into a context bundle per subagent — only the slice each subagent needs is forwarded. The Privacy Router intercepts PII before any cross-tenant or cross-vendor egress.
Evaluation
Plans are scored on three axes: completion rate, replay determinism (does re-running the same plan with the same inputs produce the same artifacts?), and human-approval acceptance rate. Continuous evaluation runs on a held-out plan set every release.