Multi-agent coordination

tl;dr

Crosslink lets you tell your agent to work on multiple features at once. Behind the scenes, agents self-coordinate through distributed locking so they never step on each other. You describe the work; the agents sort out who does what.

How it works

Lock state is stored on a crosslink/hub git branch and synchronized via git. Each agent registers a machine-local identity and claims issues through locks. Other agents see which issues are taken and work on available ones instead. All of this happens automatically – you never need to manage locks yourself.

Agent registration

You say / do:

“Work on these three features in parallel.”

You describe the features. The agent (or agents) each register unique identities before starting work.

→

Agent executes:

# Parallel session 1
crosslink agent init agent-frontend -d "Frontend specialist"
# Parallel session 2
crosslink agent init agent-backend -d "API work"
# Parallel session 3
crosslink agent init agent-infra -d "Infrastructure"

Lock coordination

You say / do:

“Go ahead, all three of you.”

You launch multiple agents. Each one automatically claims its issue via a distributed lock. No two agents will work on the same issue.

→

Agent executes:

# Each agent automatically locks its issue
crosslink session work 12   # agent-frontend claims #12
crosslink session work 15   # agent-backend claims #15
crosslink session work 18   # agent-infra claims #18

# If an agent tries a locked issue, it moves on
crosslink issue next        # skips locked issues

Lock-aware commands

These commands automatically respect lock state – no manual lock management required:

Command	Lock Behavior
`crosslink issue next`	Skips locked issues when recommending work
`crosslink session work <id>`	Warns if issue is locked by another agent
`crosslink create --work`	Only sets as active if not locked

Daemon operation

You say / do:

“Keep everything in sync while the agents work.”

The daemon runs in the background, pushing heartbeats and flushing state so agents stay coordinated without any manual intervention.

→

Agent executes:

crosslink daemon start    # background process
crosslink daemon status   # verify it's running

The daemon auto-flushes session state every 30s, pushes heartbeats every 2.5 min, and detects stale locks from dead agents.

Coordinating agents

Hub branch

Lock state lives on the crosslink/hub branch, separate from your code branches. The branch contains:

Lock files (one per locked issue)
Agent heartbeat timestamps
GPG/SSH signatures for integrity verification

Agents push to this branch when claiming/releasing locks and pull from it during crosslink sync.

Signature verification

Crosslink supports SSH signature verification for the coordination branch to ensure lock state comes from trusted agents.

You say / do:

“Make sure only our agents can modify coordination state.”

You set up trust once, and all subsequent sync operations verify signatures automatically.

→

Agent executes:

# Each agent generates a signing key on init
crosslink agent init my-agent-1

# Approve trusted agents
crosslink trust approve my-agent-1
crosslink trust list

# Sync with automatic verification
crosslink sync

Launching agents

Kickoff (single agent)

Launch a single background agent to implement a feature:

crosslink kickoff run "add batch retry logic"
crosslink kickoff run "add retry logic" --container docker  # in a container

See Kickoff for the full guide.

Swarm (multi-agent phased builds)

Coordinate multiple agents across phases from a design document:

crosslink swarm init --doc DESIGN.md
crosslink swarm launch 1

See Swarm Orchestration for the full guide.

Container execution

Run agents in isolated Docker containers instead of tmux:

crosslink kickoff run "my feature" --container docker

See Container-Based Agents for the full guide.