Roadmap

Roadmap

This roadmap keeps LightAgent focused on its core identity:

Lightweight core, composable Skills, reliable tool execution, observable traces, and OpenAI-compatible model support.

Current Foundation

Recent work has improved:

• structured run results
• structured streaming events
• catchable and formatted error codes
• tool argument validation
• memory namespace policy
• trace event export
• deterministic LightFlow workflow orchestration
• CI coverage across Python 3.10, 3.11, and 3.12
• contributor, security, and issue templates

Runtime Safety Backlog

Goal: make tool use, memory use, and final responses safer under explicit policy.

Planned work:

• input guardrails for sensitive or disallowed content
• tool-call guardrails for high-risk actions
• approval requirements for file, network, database, payment, and external-action tools
• output guardrails for JSON schema validation and PII checks
• memory guardrails for namespace isolation, provenance checks, and retrieval filtering

Expected outcome: production users can constrain actions and memory behavior without replacing LightAgent's lightweight runtime.

v0.8.0: LightFlow Workflow Orchestration

Status: released.

Goal: add deterministic workflow orchestration without turning LightAgent into a heavy framework.

Proposed shape:

flow = LightFlow()

flow.step("research", agent=research_agent)
flow.step("analyze", agent=analysis_agent, depends_on=["research"])
flow.step("write", agent=writer_agent, depends_on=["analyze"])

result = flow.run("Analyze this company")

Delivered:

• DAG-style step orchestration
• explicit step input and output passing
• step-level retry policies
• result formats: object, str, and dict
• flow traces using the same trace event system

Deferred to later versions:

• manual approval nodes
• durable execution and resume support

v0.8.5: Human in the Loop

Goal: support high-risk tasks that need explicit review before continuing.

Planned work:

• HumanApprovalTool
• approval, rejection, and edited arguments
• timeouts and rejection handling
• resume after approval

v0.9.0: Durable Execution and Resume

Goal: make long-running tasks recoverable.

Planned work:

• RunStore interface
• SQLite, Redis, and Postgres adapters
• persisted model requests, tool calls, tool results, memory reads, and memory writes
• agent.resume(run_id)
• idempotency markers for tools

v1.0.0: Stable API and Ecosystem

Goal: provide a stable, documented, tested foundation for lightweight production agents.

Planned work:

• stabilize public APIs: LightAgent, LightSwarm, LightFlow, Skill, ToolRegistry, MemoryProtocol, RunResult
• complete documentation site
• example matrix for common providers and workflows
• PyPI release automation
• benchmarks for tool-call success, multi-turn completion, cost, latency, and recovery

How to Propose Roadmap Changes

Open an issue with:

• the use case
• expected API shape
• safety impact
• testing strategy
• compatibility impact

Use existing priority labels when available: priority:p0, priority:p1, and priority:p2.