`Understudy`: teach a desktop agent by demonstrating a task once

Turns a one-off desktop workflow into a reusable skill by capturing intent, not clicks. Still macOS-only and early, but the teach-by-demo angle is fresh enough to test now.

[ KEY POINTS ]

One session can span GUI apps, browser tabs, shell tools, files, and chat, which is still rare in current agent runtimes.
The replay stores intent steps, route options, and GUI hints as fallback, so it can take faster paths instead of repeating raw macros.
Demo flow chains Google Images, Pixelmator Pro, export, and Telegram; that cross-app handoff is the real differentiator.
Install is already simple with npm install -g @understudy-ai/understudy and understudy wizard, but layers 3-4 are still partial.

Originalgithub.com/understudy-ai/understudyRead original →

// related

#0001
#0001Agents & tools MCP Servers1 hour ago
`MCP Memory Server` Adds `ToolAnnotations` Across 9 Knowledge Graph Tools
40radar
MCP Memory ServerMCP server — stores local knowledge graphs as JSONL
Agents can now distinguish read, write, and destructive graph actions before calling tools. Small change, useful for safer MCP orchestration.
- Read tools like read_graph, search_nodes, and open_nodes are marked read-only and idempotent, reducing approval friction.
- Write tools such as create_entities and add_observations are tagged non-destructive, a cleaner signal for agent planners.
- Delete tools are marked destructive and idempotent, making risk checks easier in MCP clients and IDE agents.
Source: github.com/modelcontextprotocol/servers/commit/64b1cb020Read original →
FIG-0011:1
40radar
FIG-0011:1
#0002
#0002Agents & tools Paperclip Releases5 hours ago
`Paperclip` `v2026.529.0` adds document annotations and company skills CLI
70radar
PaperclipCoding agent platform — document and skill workflows
Agent work now gets passage-level review threads and managed skill catalogs. Useful if you run self-hosted agent workflows; otherwise it is a strong product-direction signal.
- Issue documents now support revision-aware inline annotation threads with stable anchors, reducing vague review comments around agent-written specs.
- New skills CLI covers install, reset, audit, export, and assignment; skill provenance is tracked across bundled, catalog, runtime, and adapter sources.
- Fresh private self-hosted installs get a one-time first-admin browser claim, removing the invite bootstrap problem on LAN deployments.
- Claude Local can refresh its Anthropic model catalog from the UI, so new Claude models can appear without waiting for a Paperclip release.
Source: github.com/paperclipai/paperclip/releases/tag/v2026.529.Read original →
FIG-0021:1
70radar
FIG-0021:1
#0003
#0003Agents & tools Claude Code Releases17 hours ago
`Claude Code` `v2.1.158` adds Auto mode on Bedrock, Vertex, and Foundry
60radar
Claude CodeCoding agent CLI — automates code work with Claude
Auto mode now reaches managed-cloud deployments for Opus 4.7 and Opus 4.8. Useful only if you run through Bedrock, Vertex, or Foundry.
- CLAUDE_CODE_ENABLE_AUTO_MODE=1 enables the feature, so rollout is opt-in rather than default behavior.
- Support covers Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8; direct API users get no new signal here.
- Practical value is higher for teams keeping LLM traffic inside cloud-vendor rails; local CLI-only workflows can skip this update.
Source: github.com/anthropics/claude-code/releases/tag/v2.1.158Read original →
FIG-0031:1
60radar
FIG-0031:1

`Understudy`: teach a desktop agent by demonstrating a task once

// related

`MCP Memory Server` Adds `ToolAnnotations` Across 9 Knowledge Graph Tools

`Paperclip` `v2026.529.0` adds document annotations and company skills CLI

`Claude Code` `v2.1.158` adds Auto mode on Bedrock, Vertex, and Foundry