beads/docs/MOLECULES.md

Molecules: Work Graphs in Beads

This doc explains how beads structures and executes work. Start here if you're building workflows.

TL;DR

1. Work = issues with dependencies. That's it. No special types needed.
2. Dependencies control execution. blocks = sequential. No dep = parallel.
3. Molecules are just epics. Any epic with children is a molecule. Templates are optional.
4. Bonding = adding dependencies. bd mol bond A B creates a dependency between work graphs.
5. Agents execute until blocked. When all ready work is done, the workflow is complete.

The Execution Model

How Work Flows

An agent picks up a molecule (epic with children). They execute ready children in parallel until everything closes:

epic-root (assigned to agent)
├── child.1 (no deps → ready)      ← execute in parallel
├── child.2 (no deps → ready)      ← execute in parallel
├── child.3 (needs child.1) → blocked until child.1 closes
└── child.4 (needs child.2, child.3) → blocked until both close

Ready work: bd ready shows issues with no open blockers. Blocked work: bd blocked shows what's waiting.

Dependency Types That Block

Type	Semantics	Use Case
blocks	B can't start until A closes	Sequencing work
parent-child	If parent blocked, children blocked	Hierarchy (children parallel by default)
conditional-blocks	B runs only if A fails	Error handling paths
waits-for	B waits for all of A's children	Fanout gates

Non-blocking types: related, discovered-from, replies-to - these link issues without affecting execution.

Default Parallelism

Children are parallel by default. Only explicit dependencies create sequence:

# These three tasks run in PARALLEL (no deps between them)
bd create "Task A" -t task
bd create "Task B" -t task
bd create "Task C" -t task

# Add dependency to make B wait for A
bd dep add <B-id> <A-id>   # B depends on A (B needs A)

Multi-Day Execution

An agent works through a molecule by:

1. Getting ready work (bd ready)
2. Claiming it (bd update <id> --status in_progress)
3. Doing the work
4. Closing it (bd close <id>)
5. Repeat until molecule is done

If the molecule is blocked by another molecule:

• Agent either waits, or
• Agent continues into the blocking molecule (compound execution)

Bonding enables compound execution: When you bond molecule A to molecule B, the agent can traverse both as one logical unit of work.

Molecules vs Epics

They're the same thing. A molecule is just an epic (parent + children) with workflow semantics.

Term	Meaning	When to Use
Epic	Parent issue with children	General term for hierarchical work
Molecule	Epic with execution intent	When discussing workflow traversal
Proto	Epic with template label	Reusable pattern (optional)

You can create molecules without protos - just create an epic and add children:

bd create "Feature X" -t epic
bd create "Design" -t task --parent <epic-id>
bd create "Implement" -t task --parent <epic-id>
bd create "Test" -t task --parent <epic-id>
bd dep add <implement-id> <design-id>   # implement needs design
bd dep add <test-id> <implement-id>      # test needs implement

Bonding: Connecting Work Graphs

Bond = create a dependency between two work graphs.

bd mol bond A B                    # B depends on A (sequential by default)
bd mol bond A B --type parallel    # Organizational link, no blocking
bd mol bond A B --type conditional # B runs only if A fails

What Bonding Does

Operands	What Happens
epic + epic	Creates dependency edge between them
proto + epic	Spawns proto as new issues, attaches to epic
proto + proto	Creates compound template

The Key Insight

Bonding lets agents traverse compound workflows. When A blocks B:

• Completing A unblocks B
• Agent can continue from A into B seamlessly
• The compound work graph can span days

This is how Gas Town runs autonomous workflows - agents follow the dependency graph, handing off between sessions, until all work closes.

Phase Metaphor (Templates)

For reusable workflows, beads uses a chemistry metaphor:

Phase	Name	Storage	Synced	Purpose
Solid	Proto	.beads/	Yes	Frozen template
Liquid	Mol	.beads/	Yes	Active persistent work
Vapor	Wisp	.beads/ (Wisp=true)	No	Ephemeral operations

Phase Commands

bd pour <proto>                  # Proto → Mol (persistent instance)
bd ephemeral create <proto>           # Proto → Wisp (ephemeral instance)
bd mol squash <id>               # Mol/Wisp → Digest (permanent record)
bd mol burn <id>                 # Wisp → nothing (discard)

When to Use Each Phase

Use Case	Phase	Why
Feature work	Mol (pour)	Persists across sessions, audit trail
Patrol cycles	Wisp	Routine, no audit value
One-shot ops	Wisp	Scaffolding, not the work itself
Important discovery during wisp	Mol (--pour)	"This matters, save it"

Common Patterns

Sequential Pipeline

bd create "Pipeline" -t epic
bd create "Step 1" -t task --parent <pipeline>
bd create "Step 2" -t task --parent <pipeline>
bd create "Step 3" -t task --parent <pipeline>
bd dep add <step2> <step1>
bd dep add <step3> <step2>

Parallel Fanout with Gate

bd create "Process files" -t epic
bd create "File A" -t task --parent <epic>
bd create "File B" -t task --parent <epic>
bd create "File C" -t task --parent <epic>
bd create "Aggregate" -t task --parent <epic>
# Aggregate needs all three (waits-for gate)
bd dep add <aggregate> <fileA> --type waits-for

Dynamic Bonding (Christmas Ornament)

When the number of children isn't known until runtime:

# In a survey step, discover polecats and bond arms dynamically
for polecat in $(gt polecat list); do
  bd mol bond mol-polecat-arm $PATROL_ID --ref arm-$polecat --var name=$polecat
done

Creates:

patrol-x7k (wisp)
├── preflight
├── survey-workers
│   ├── patrol-x7k.arm-ace (dynamically bonded)
│   ├── patrol-x7k.arm-nux (dynamically bonded)
│   └── patrol-x7k.arm-toast (dynamically bonded)
└── aggregate (waits for all arms)

Agent Pitfalls

1. Temporal Language Inverts Dependencies

Wrong: "Phase 1 comes before Phase 2" → bd dep add phase1 phase2 Right: "Phase 2 needs Phase 1" → bd dep add phase2 phase1

Use requirement language. Verify with bd blocked.

2. Assuming Order = Sequence

Numbered steps don't create sequence. Dependencies do:

# These run in PARALLEL despite names
bd create "Step 1" ...
bd create "Step 2" ...
bd create "Step 3" ...

# Add deps to sequence them
bd dep add step2 step1
bd dep add step3 step2

3. Forgetting to Close Work

Blocked issues stay blocked forever if their blockers aren't closed. Always close completed work:

bd close <id> --reason "Done"

4. Orphaned Wisps

Wisps accumulate if not squashed/burned:

bd ephemeral list           # Check for orphans
bd mol squash <id>     # Create digest
bd mol burn <id>       # Or discard
bd ephemeral gc             # Garbage collect old wisps

Layer Cake Architecture

For reference, here's how the layers stack:

Formulas (JSON compile-time macros)      ← optional, for complex composition
    ↓
Protos (template issues)                  ← optional, for reusable patterns
    ↓
Molecules (bond, squash, burn)            ← workflow operations
    ↓
Epics (parent-child, dependencies)        ← DATA PLANE (the core)
    ↓
Issues (JSONL, git-backed)                ← STORAGE

Most users only need the bottom two layers. Protos and formulas are for reusable patterns and complex composition.

Commands Quick Reference

Execution

bd ready                         # What's ready to work
bd blocked                       # What's blocked
bd update <id> --status in_progress
bd close <id>

Dependencies

bd dep add <issue> <depends-on>  # issue needs depends-on
bd dep tree <id>                 # Show dependency tree

Molecules

bd pour <proto> --var k=v        # Template → persistent mol
bd ephemeral create <proto>           # Template → ephemeral wisp
bd mol bond A B                  # Connect work graphs
bd mol squash <id>               # Compress to digest
bd mol burn <id>                 # Discard without record

Related Docs

• CLI_REFERENCE.md - Full command reference
• ARCHITECTURE.md - System internals
• ../CLAUDE.md - Quick agent reference