This was a big execution-focused release. Orchestration is more capable, planning chat gained practical tools, skills became more reliable, and Telegram got better at keeping you in the loop when agents need input.
ORCHESTRATE: Orchestration For Any Planning Entity
Orchestration is no longer a special case. It now works as a general execution mode that can coordinate work across your plan hierarchy.
- Orchestrate any entity type - Not just leaf tasks. Objectives, projects, and processes can now act as orchestration roots.
- Better runnable logic - Execution is more deliberate about prerequisites (including previous-sibling completion), reducing out-of-order work.
- Automatic activation of draft ancestors - When you submit work, ancestor entities are activated as needed so scheduling is not blocked by a forgotten draft status.
- Recurring work stays alive - Recurring processes can remain active across runs, and completed children can be re-activated on the next recurrence.
Result: fewer stuck plans, fewer manual status flips, and better end-to-end momentum on long-running efforts.
Planning Chat: Bash + File Tools
Planning chat is now more than "talk about doing the work". Agents can execute and report real work through built-in tools.
- Bash execution - Run shell commands as part of a step (useful for quick checks, scripts, and automation)
- File operations - Read/write/update files as part of tool-driven execution
- Clearer rendering - Tool outputs are rendered in a way that is easier to read and audit
- Fewer streaming crashes - More defensive handling of malformed/partial streamed message parts
If you use Kindship to actually ship changes (not just brainstorm), this improves the reliability of the execution loop dramatically.
Skills: Now Live In The Container Workspace
Default skills moved out of the GitHub documents repo and into the agent container, where they are more reliable and less intrusive.
- Skills installed directly into the agent container during provisioning
- Embedded defaults - Skills are embedded at build time for worker/serverless environments, reducing runtime dependency issues
- Cleaner documents experience - The Documents tab hides the internal
skills/folder by default so your actual work is front-and-center
Result: fewer sync headaches, fewer repo artifacts, and more consistent agent behavior.
Account-Level Credentials (Shared Across Agents)
Instead of configuring AI credentials per agent (or per container), you can now manage shared credentials at the account level.
- Works with Claude Code and Gemini CLI - Configure shared credentials once for the account
- One configuration, many agents - Set credentials once and apply them across all agents in the account
- Status visibility - See whether credentials are configured and whether container CLIs are authenticated/installed
- Safer storage - Additional hardening for credential storage and secret handling
This is especially valuable for team accounts where you want consistent tooling across multiple agents.
Activity: Better Visibility Into Runs
The Activity experience is more useful when you are actively executing plans.
- Runs are visible - Execution runs now show up in the Activity tab
- Better filtering - Run filtering is more accurate, so you see what matters
- Cleaner UI - Small layout tweaks reduce noise (for example, the polling spinner presentation)
Ask-User Prompts Via Telegram (Execution Keeps Moving)
When a plan hits an ASK_USER step, agents can now reach you on Telegram instead of waiting silently in the web UI.
- Proactive Telegram prompt - You get the question where you are, without babysitting the browser
- Auto-switch to the asking agent - Telegram sessions switch to the correct agent so your reply lands in the right place
- Reply directly in Telegram - Answer inline to unblock the run
Telegram Controls: /tool and /agents
Two quality-of-life improvements for Telegram power users:
/tooltoggle - Turn tool-call notification messages on/off in Telegram/agentsalias - An additional command that maps to/agentfor faster switching
Reliability Improvements Across The Board
A grab bag of fixes that reduce friction when you are running agents at scale:
- More reliable CLI login - Browser authentication works more consistently across environments
- More robust provisioning - Better logging, fewer silent failures when workflows fail to start
- Container manager stability - More correct handling of Docker stream frames (fewer broken logs/terminals)
- Less log spam - Noisy "no tasks available" and similar messages were reduced so real issues stand out
- Security hardening - Stronger path sanitization and safer credential storage to reduce foot-guns