Orchestra AI by Hyperdrift · Field notes
The AI-native
organisation.
Practical notes from building and running our apps. Find a working example, inspect the evidence, and see where it could help you.
- 01Read →
Delegation / Helm
AI agent permissions: how much authority should you delegate?
Helm separates diagnosis from action and restricts disruptive operations to Cargo. Its completion-reporting gap shows why permissions and outcome checks need separate tests.
Recorded demonstration - 02Read →
Priorities / Standup
AI daily briefs: turn activity into useful priorities
Standup puts people waiting for a maintainer before routine work. Its recorded output and later MCP checks show why each recommendation needs a source the reader can inspect.
Public project - 03Read →
AI engineering / Databases and agent state
Lakebase vs MongoDB and Cosmos DB: which approach fits?
Compare Lakebase’s Postgres approach with MongoDB and Cosmos DB document models: relationships, transactions, use cases and when Databricks integration matters.
Illustrative decision guide - 04Read →
AI engineering / Streaming and event contracts
LangChain vs Databricks AppKit for agent streaming
LangChain supplies agent updates; Databricks AppKit adds an integrated app server and agent host. Compare streaming, SSE, approval and recovery responsibilities.
Illustrative architecture guide - 05Read →
Conversation / Hyperdrift First Officer
The dashboard that speaks first
In a recorded Cargo recovery, a spoken instruction starts the work. First Officer checks the response and reports back, with the evidence still on screen.
Recorded sandbox recovery · 29 September 2026 - 06Read →
History and human capability / UI and accessibility
Is UI holding us back?
New interfaces can remove work or move it into checking and correction. A historical argument for judging the whole task.
Historical synthesis - 07Read →
Accessibility today / UI and accessibility
Web accessibility: can people finish the job?
An accessible button is useful. An accessible way to finish is the point. What the current evidence tells us about agents and access.
Research and proposed evaluation - 08Read →
WebMCP and shared state / UI and accessibility
Stop making agents hunt for buttons
WebMCP lets a page expose actions to a browser agent. Radar shows how that can keep the person and agent working in the same place.
Radar implementation and draft specification - 09Read →
Traction / First Officer / UI and accessibility
A spoken promise is not a completed action
A Cargo recovery request becomes an action, a checked result and a spoken report. What this bounded demonstration can teach app designers.
Recorded sandbox demonstration - 10Read →
Voice through MCP / UI and accessibility
Say the job once. Keep control of the result.
Dictating to an assistant can become a way to direct connected tools. Our proposed experiment starts with a scoped read and a draft.
Proposed experiment - 11Read →
The future of UI / UI and accessibility
Your work should survive a change of interface
Speak, inspect, correct and continue. A future interface should preserve the task when the person changes how they interact.
Design position and research agenda - 12Read →
Security maintenance / Orchestra OS
Should a security fix wait for the next feature release?
A finished app still needs repairs. Patch it from production, carry the fix back to development, and refuse any later release that would undo it.
Proposed reference model - 13Read →
App operations / Orchestra OS
Your app works. Who runs it now?
A live app creates five jobs every week. Give them named owners, agents with limits and receipts, and keep the decisions. What that looks like, and what to ask whoever runs your app.
Our own apps run on it first - 14Read →
A peek into Traction / Traction
Here’s the new deal. UI is for intel. Agents are for operating.
A peek into Traction’s 24 October release: the board shows what the Growth Manager and Builder are doing and what they delivered. The agents do the work. We decide what becomes work.
Our Orchestra board · 24 October 2026 - 18Read →
OSSAgent telemetry / OpenTelemetry GenAI conventions
We got our first PR merged into OpenTelemetry
OpenTelemetry's GenAI conventions now say how an agent framework can share its conversation ID with the model and tool calls nested inside it. Review narrowed the rule to that one ID; agent naming continues in a separate draft.
Merged upstream · 30 October 2026