Plans
Turn a work objective into an explicit plan with ordered steps, dependencies, expected artifacts, and review boundaries.
LexRunner is the flow layer for multi-step and multi-agent work. It consumes memory, capability, behavior, and authority surfaces, then turns a plan into gated execution with receipts.
The next important direction is contained swarm work: separate execution roots where agents can explore in parallel without leaking state, credentials, or half-finished changes into each other.
Lex -> memory and receipts
AXF -> workspace capabilities
LexSona -> behavioral constraints
LexThority -> authority decisions
LexRunner -> flow, gates, isolation, merge
Turn a work objective into an explicit plan with ordered steps, dependencies, expected artifacts, and review boundaries.
Run the right checks before a change advances: tests, lint, policy checks, authority checks, review hooks, and human escalation.
Capture what happened at each step so another agent or human can understand the state without reconstructing the whole run.
Give agents a bounded place to work. The design goal is safer than raw worktrees: scoped roots, explicit mounted state, and clear promotion.
Fan out independent work and merge it back through deterministic decisions, not conversational handoffs.
When a tool, permission, test, or policy blocks work, return an actionable explanation instead of continuing blindly.
Agent swarm behavior should not mean every agent shares the same checkout, credentials, and context. The useful version is bounded parallelism.
Each agent receives a contained working area with an explicit relationship to the project root and a known promotion path.
A PM-shaped agent, analyst-shaped agent, developer-shaped agent, and operator-shaped agent should not all carry the same authority posture.
LexRunner should not become an IAM broker. It should discover when the environment cannot perform an effect and report the exact missing operator action.
Artifacts move from contained work into canonical branches, queues, tickets, or reports only after gates and receipts make the decision reviewable.
| Not memory | Lex owns durable memory, Frames, policy context, and recall. |
|---|---|
| Not a persona engine | LexSona owns behavioral constraints and posture. |
| Not an authority broker | LexThority owns effect decisions; enterprise IAM remains the source of truth for actual credentials. |
| Not a command catalog | AXF owns workspace capability declarations and adapter contracts. |
| Not a rigid SOP engine | Runner should leave room for agent reasoning while bounding effects through gates, receipts, and promotion rules. |
LexRunner is not published on npm and the source is controlled-access while the containment and swarm model is still settling.