Skip to content

Working with agents

Run focused sessions in a shared project, coordinate edits with file locks, and pass clear work between agents.

Parallel sessions

Each session keeps its own conversation, model selection, and task. Sessions in the same project work on the same checkout. Changes become visible to the other sessions without creating an extra worktree for every agent.

File locks

Supported Koyd agent tools coordinate file access through the server’s lock system. When another session owns an incompatible lock, the conflicting operation is blocked or waits through the supported coordination flow. Agents can choose other work while a file is occupied.

Inspect the file-lock indicator to see who holds a file. Swarm tasks also use declared file footprints and leases to coordinate work. Keeping tasks’ write scopes clear makes both mechanisms more useful.

Good to know · Locks are coordination, not an operating-system sandbox. An external editor, an unintegrated tool, or an opaque command can still change a file. Do not treat shared-checkout locks as branch isolation or a guarantee against every conflict.

Choose a mode

Modes shape the instructions and capabilities used for a session. Select a mode suitable for investigation, planning, implementation, or review. Check the mode’s configured tools before giving it access to important files.

Save reusable agents

The Lab holds agents, modes, prompt templates, skills, MCP connections, and swarm presets. Save the instructions and tool choices you want to reuse. Customization lives on the selected instance, so another server can have a different setup.

Delegate within a turn

A session can delegate bounded work to subagents when the selected agent tools support it. Give children a specific objective, the context they need, and a clear expected result. Koyd integrates supported child sessions with its parent identity, file locks, and configured MCP access.

Good to know · Subagents consume provider usage too. A larger team is useful only when the task can benefit from parallel work.

Hand off work

Pass useful context to another session through the handoff flow. Select the relevant response or prepare a concise brief, pick the destination, and review what will be sent. Dependency chains can sequence work so the next task begins when its prerequisite is ready.

Queued prompts and dependency handling run on the server. Closing the PWA does not pause a handoff or its downstream work.

Skills and MCP tools

Skills provide reusable guidance; MCP connections provide tools from configured services. Manage them in the Lab and select the resources appropriate to the task. Review tool access and the credentials you supply before enabling an integration.