Why Harness
Many agent stacks work well for personal assistants, but hit real limits in production:- Distributed workspace state: local directory assumptions break in multi-replica deployments.
- Execution safety: user-driven shell/code execution must be isolated from host processes.
- Storage abstraction: agent logic should not be coupled to one storage store.
- Subagent orchestration: task delegation and lifecycle management become complex quickly.
- Memory governance: context growth and cross-session fact retention need first-class support.
Core Design
Workspace as source of truth
Harness organizes long-lived agent state under a structured workspace:- persona and rules (
AGENTS.md) - long-term memory (
MEMORY.md) - domain knowledge
- skills
- subagent definitions
- session artifacts
After each call, new facts and state are persisted back.
AbstractFilesystem as portability layer
Harness decouples logical file operations (read, write, ls, grep) from physical storage/execution:
- local disk
- remote storage
- sandbox filesystem
- composite routing across stores
What 1.1 Delivers
- Workspace-driven runtime for durable identity and evolving behavior
- Pluggable filesystem abstraction for local/remote/sandbox execution
- Built-in context governance (compaction + layered memory)
- Subagent orchestration with sync/async delegation and task lifecycle
Typical Deployment Scenarios
Personal productivity agents
- local execution
- durable memory and persona
- workspace-driven skill iteration
Enterprise data/service agents
- sandboxed execution for untrusted input
- distributed memory/session continuity across replicas
- async subagent orchestration for long-running tasks
API-first online agents
- strict tool boundaries (no implicit shell exposure)
- shared remote state for cross-instance continuity
- stable multi-turn behavior under production traffic