Selectools Development Roadmap

June 15, 2026 · View on GitHub

An open-source project from NichevLabs.

Status Legend

✅ Implemented - Merged and available in latest release

🔵 In Progress - Actively being worked on

🟡 Planned - Scheduled for implementation

⏸️ Deferred - Postponed to later release

❌ Cancelled - No longer planned

v0.17.0 ✅ Eval Framework 39 evaluators → A/B testing → snapshots → regression → HTML/JUnit → CI → templates

v0.17.1 ✅ MCP Client/Server MCPClient → mcp_tools() → MCPServer → MultiMCPClient → tool interop

v0.17.3 ✅ Agent Runtime Controls Token budget → Cancellation → Cost attribution → Structured results → Approval gate → SimpleStepObserver

v0.17.4 ✅ Agent Intelligence Token estimation → Model switching → Knowledge memory enhancement (4 store backends)

v0.17.5 ✅ Bug Hunt & Async Guardrails 91 validated fixes (13 critical, 26 high, 52 medium+low) → Async guardrails → 40 regression tests → 5 new Common Pitfalls

v0.17.6 ✅ Quick Wins ReAct/CoT reasoning strategies → Tool result caching → Python 3.9–3.13 CI matrix

v0.17.7 ✅ Caching & Context Semantic caching → Prompt compression → Conversation branching (55 tests, 3 examples)

v0.18.0 ✅ Multi-Agent Orchestration + Composable Pipelines AgentGraph → GraphState → Typed reducers → Resume-from-yield interrupts → Scatter fan-out → Checkpointing → SupervisorAgent → Graph visualization → Pipeline → @step → | operator → parallel() → branch()

v0.19.0 ✅ Serve, Deploy & Complete Composition selectools serve CLI → Playground UI → YAML config → 5 agent templates → Structured AgentConfig → compose() → retry() / cache_step() → Type-safe step contracts → Streaming composition → pipeline.astream() → PostgresCheckpointStore → TraceStore (3 backends) → selectools doctor

v0.19.1 ✅ Advanced Agent Patterns PlanAndExecute → ReflectiveAgent → Debate → TeamLead → 50+ evaluators

v0.19.2 ✅ Enterprise Hardening Security audit → Stability markers (@stable/@beta/@deprecated) → Deprecation policy → Compatibility matrix → trace_to_html() waterfall viewer → SBOM → Property-based tests (Hypothesis) → Concurrency smoke suite → 5 production simulations → 3,344 tests, 76 examples

v0.19.3 ✅ Stability Markers Applied to All Public APIs @stable on 60+ core symbols → @beta on 30+ orchestration/pipeline/patterns symbols → Full stability introspection via .stability on every exported class and function

v0.20.0 ✅ Visual Agent Builder Zero-install web UI → Drag-drop graph builder → YAML/Python export → Self-contained HTML (no React, no CDN) → One command: selectools serve --builder

v0.20.1 ✅ Builder Polish + Starlette + GitHub Pages UI polish (20 features) → _static/ architecture split → Starlette ASGI app → Serverless mode (client-side AI/runs) → GitHub Pages deployment → Design system

v0.21.0 ✅ Connector Expansion + Multimodal + Observability FAISS → Qdrant → pgvector vector stores → Azure OpenAI provider → Multimodal messages (images, audio) → CSV/JSON/HTML/URL document loaders → OTel observer → Langfuse observer → Code execution, web search, GitHub, DB toolbox tools

v0.22.0 ✅ Competitor-Informed Bug Fixes + Loop Detection + Ruff Tooling 38 bug fixes from 3 rounds mining Agno/PraisonAI/LangChain/LangGraph/ CrewAI/n8n/LlamaIndex/AutoGen/LiteLLM/Pydantic-AI/Haystack (~325k stars) → Loop detection (Repeat, Stall, PingPong) with RAISE / INJECT_MESSAGE policies → Dev tooling consolidated: Black + isort + flake8 → Ruff → 30-recipe cookbook expansion → 95 runnable examples

v0.23.0 ✅ Supabase Sessions + Builder RAG SupabaseSessionStore → 4th SessionStore backend (JSON/SQLite/Redis/Supabase) → Visual builder: first-class Retriever (RAG) + Session Store node types → 7 vector-store backends in builder (memory/SQLite/Chroma/Pinecone/FAISS/Qdrant/pgvector) → Hybrid (BM25 + vector + RRF) + cross-encoder rerank toggles → New presets: Hybrid RAG, Multi-Tenant RAG (pgvector + Supabase session) → 8 post-ship code-gen fixes in builder (embedder class names, HybridSearcher params, etc.) → 96 runnable examples, 5332 tests total

v0.24.0 ✅ Production Interop Agent-as-API (AgentAPI: REST + SSE + session CRUD + auth) → A2A protocol (Agent Card + JSON-RPC 2.0 server/client) → LiteLLMProvider (100+ models) → RouterProvider (cost-optimized tier routing) → Anthropic prompt caching → UnifiedMemory (conversation/knowledge/entity/episodic tiers) → Cross-session search on all 4 SessionStore backends → KnowledgeBackend (Supabase/Redis) → ToolResult base + Artifact side-channel → Deferred confirmation flow (selectools.pending) → Toolbox expansion: 15 new tools (33 → 48) → Gemini schema sanitization + flash-lite compat → 106 runnable examples, 5968 tests total

v0.25.0 ✅ Hardening & v1.0 Prep Planning-as-config (AgentConfig(planning=...)) → Agent-level HITL (ToolConfig(require_approval=...)) → Tool result compression → Knowledge pre-save sanitizers → Pending intent hooks (pop_if_intent, tighten_ttl) → Stability marking sweep: 433 public symbols 100% marked (205 stable / 228 beta), 19 beta→stable promotions, stability on all 123 public modules, CI gate → Wart removal: clone_for_isolation() public, all reconciled (+11 exports), AgentConfig.hooks REMOVED (BREAKING) → Security audit published (docs/SECURITY_AUDIT.md) → 0.x→1.0 migration guide → Compatibility matrix refresh → 111 runnable examples, 7268 tests total

v0.26.0 ✅ Safety Patch & Verified Registry Confirm-parser negation veto (non-leading negation no longer fired destructive CONFIRM) → Model registry refresh: 152 → 115, every entry source-verified, opus-4-1 pricing corrected, retired-model constants REMOVED (BREAKING) → Cache-aware calculate_cost → A2A -32602 on malformed parts → Gemini embedding dimension constant 3072 → 111 runnable examples, 7420 tests total

v1.0.0 🟡 Stable Release (bake window — code-complete) API freeze ✅ (warts removed in v0.25) → Stability markers on all modules ✅ → Security audit published ✅ → Compatibility matrix ✅ → 0.x→1.0 migration guide ✅ → Deprecation policy → Remaining at tag time: drop Python 3.9 → PyPI classifier: Production/Stable

v0.19.1: Advanced Agent Patterns ✅

Higher-level agent architectures built on the v0.18.0 orchestration primitives. Closes the "Advanced patterns" competitive gap. Each pattern is a standalone class — they wire up the AgentGraph topology for you.

PlanAndExecute Agent

from selectools.patterns import PlanAndExecuteAgent

agent = PlanAndExecuteAgent(
    planner=planner_agent,
    executors={"research": researcher, "write": writer, "review": reviewer},
)
result = agent.run("Write a technical blog post about vector databases")
# Planner creates structured plan → executors handle each step → result aggregated

ReflectiveAgent

from selectools.patterns import ReflectiveAgent

agent = ReflectiveAgent(
    actor=writer_agent,
    critic=reviewer_agent,
    max_reflections=3,
    stop_condition="approved",
)
result = agent.run("Draft a press release")
# Actor produces draft → Critic evaluates → Actor revises → repeat until approved

Debate Pattern

from selectools.patterns import DebateAgent

agent = DebateAgent(
    agents={"optimist": optimist_agent, "skeptic": skeptic_agent},
    judge=judge_agent,
    max_rounds=3,
)
result = agent.run("Should we adopt microservices?")
# Agents argue positions → Judge synthesizes final answer

TeamLead Pattern

from selectools.patterns import TeamLeadAgent

agent = TeamLeadAgent(
    lead=lead_agent,
    team={"analyst": analyst, "engineer": engineer, "writer": writer},
    delegation_strategy="dynamic",  # or "sequential", "parallel"
)
result = agent.run("Investigate and fix the billing discrepancy")
# Lead delegates tasks, reviews work, coordinates handoffs

Expanded Eval Suite (50 evaluators, up from 39)

11 new evaluators across two categories:

New deterministic (+8): ReadabilityEvaluator, AgentTrajectoryEvaluator, ToolEfficiencyEvaluator, SemanticSimilarityEvaluator, MultiTurnCoherenceEvaluator, JsonSchemaEvaluator, KeywordDensityEvaluator, ForbiddenWordsEvaluator

New LLM-as-judge (+4): FactConsistencyEvaluator, CustomRubricEvaluator, AnswerAttributionEvaluator, StepReasoningEvaluator

Feature	Status	Impact	Effort
PlanAndExecute	✅	High	Medium
ReflectiveAgent	✅	High	Medium
Debate	✅	Medium	Medium
TeamLead	✅	Medium	Medium
50 evaluators	✅	High	Medium

Quality Infrastructure

Ralph loop — autonomous hunt-and-fix convergence system (scripts/ralph_bug_hunt.sh, /ralph-bug-hunt skill)
Bandit in CI — security scan job on every push
Property-based tests — Hypothesis suite for structural invariants
Thread-safety smoke suite — 10-thread × 20-op concurrency tests
Production simulations — 16 integration tests covering memory pressure, provider failover, tool errors, concurrent load

v0.19.2: Enterprise Hardening ✅

Focus: Production readiness and developer trust signals before the Visual Agent Builder in v0.20.0.

Enterprise Hardening

Feature	Status	Impact	Effort
Security audit (bandit + manual nosec review)	✅	High	Medium
Stability markers (`@stable`, `@beta`, `@deprecated`)	✅	Medium	Small
Deprecation policy (2-version window, `docs/DEPRECATION_POLICY.md`)	✅	Medium	Small
Compatibility matrix (Python × provider SDK × optional deps)	✅	Medium	Small
SBOM (`sbom.json` via CycloneDX, published in repo)	✅	Low	Small
Enhanced trace viewer (`trace_to_html()` waterfall HTML)	✅	High	Medium

Quality Infrastructure

Feature	Status	Impact	Effort
Property-based tests (Hypothesis)	✅	High	Medium
Thread-safety smoke suite	✅	High	Medium
Production simulations (5 new)	✅	High	Medium

v0.19.3: Stability Markers Applied ✅

Focus: Apply @stable and @beta markers to every public symbol in the library, completing the stability annotation work started in v0.19.2.

Stable APIs (60+ symbols)

Core types, providers, agent, memory, tools, evals, guardrails, sessions, knowledge, cache, cancellation, token estimation, analytics, audit — all marked @stable. Breaking changes to these require a major version bump.

Beta APIs (30+ symbols)

Orchestration (AgentGraph, SupervisorAgent), pipelines (Pipeline, @step, parallel, branch), patterns (PlanAndExecuteAgent, ReflectiveAgent, DebateAgent, TeamLeadAgent), and composition (compose) — marked @beta. These may change in a minor release.

Introspection

from selectools import Agent, AgentGraph, PlanAndExecuteAgent
print(Agent.__stability__)              # "stable"
print(AgentGraph.__stability__)         # "beta"
print(PlanAndExecuteAgent.__stability__)  # "beta"

v0.20.0: Visual Agent Builder ✅

The headline feature: a zero-install web UI for designing, testing, and exporting agent configurations. Served by selectools serve --builder — no separate app, no subscription, no desktop install required.

Why a dedicated release: LangGraph Studio is a paid desktop app. AutoGen Studio is a separate project. selectools ships a full visual builder in one command. This deserves its own announcement.

A web-based UI for designing, testing, and exporting agent configurations. Zero-install — served by selectools serve --builder.

┌─────────────────────────────────────────────────────┐
│  Visual Agent Builder                    [Export]   │
├─────────────┬───────────────────────────────────────┤
│             │                                       │
│  Components │    ┌──────────┐    ┌──────────┐       │
│  ─────────  │    │ Planner  │───▶│  Writer  │       │
│  ☐ Agent    │    └──────────┘    └────┬─────┘       │
│  ☐ Tool     │                         │             │
│  ☐ Router   │                    ┌────▼─────┐       │
│  ☐ Gate     │                    │ Reviewer │       │
│  ☐ Parallel │                    └──────────┘       │
│             │                                       │
├─────────────┼───────────────────────────────────────┤
│  Properties │    Model: gpt-4o  │ Tools: 3          │
│  ─────────  │    Strategy: plan │ Budget: \$0.50     │
│  Name: ...  │                                       │
│  Model: ... │    [▶ Test Run]   [💾 Save YAML]      │
└─────────────┴───────────────────────────────────────┘

Features:

Drag-and-drop graph builder for AgentGraph topologies
Node palette: Agent, Tool, Router (conditional), Gate (HITL), Parallel group
Visual edge wiring with routing condition editor
Per-node configuration panel (model, tools, system prompt, budget)
Live test: run the graph against real providers from the UI
Export: generates agent.yaml or Python code
Import: load existing YAML configs into the builder
Served by selectools: selectools serve --builder (zero frontend deps)
Built as self-contained HTML/JS (same pattern as playground.py)

Technical approach:

Single HTML file with embedded JS (no React, no build step)
Canvas-based graph rendering (or SVG with drag handlers)
Backend: new /builder endpoint on AgentServer
- GET /builder — serves the HTML
- POST /builder/validate — validates graph structure
- POST /builder/export — generates YAML or Python
- POST /builder/run — executes the designed graph
State stored in browser localStorage (no server state)

Why this matters:

LangGraph has LangGraph Studio (paid, desktop app)
CrewAI has no visual builder
AutoGen has AutoGen Studio (separate app)
selectools: zero-install, runs in browser, exports to YAML/Python

Feature	Status	Impact	Effort
Graph canvas (drag-drop nodes + edges)	✅	High	Large
Node configuration panel	✅	High	Medium
YAML export/import	✅	High	Small
Python code export	✅	Medium	Small
Live test execution	✅	High	Medium
Self-contained HTML (no build step)	✅	High	Medium

v0.20.1: Builder Polish + Starlette + GitHub Pages ✅

UI polish (20 features), _static/ architecture split, Starlette ASGI app, serverless mode (client-side AI/runs), GitHub Pages deployment, design system.

Visual builder live at: https://selectools.dev/builder/
Examples gallery: https://selectools.dev/examples/
4,612 tests (95% coverage), 76 examples, 50 evaluators, 152 models

Feature	Status	Impact	Effort
UI polish (20 features)	✅	High	Medium
_static/ architecture split	✅	Medium	Small
Starlette ASGI app	✅	High	Medium
Serverless mode (client-side AI/runs)	✅	High	Medium
GitHub Pages deployment	✅	High	Small
Design system	✅	Medium	Small
Eval badges on builder nodes	✅	Medium	Small

v0.21.0: Connector Expansion + Multimodal + Observability ✅

Shipped: FAISS + Qdrant + pgvector vector stores, CSV/JSON/HTML/URL document loaders, Azure OpenAI provider, OpenTelemetry + Langfuse observers, multimodal ContentPart + image_message() across OpenAI/Anthropic/Gemini/Ollama, new code/search/github/db toolbox modules (9 tools). 5215 tests (95% coverage), 88 examples, 5 LLM providers, 7 vector stores, 152 models.

Close integration gaps, add multimodal support (images/audio), and ship enterprise-grade observability (OTel + Langfuse). Full spec: .private/07-v0.21.0-connector-expansion.md

Current Inventory

Category	Count	Items
Document Loaders	4	text, file, directory, PDF
Vector Stores	4	Memory, SQLite, Chroma, Pinecone
Embedding Providers	4	OpenAI, Anthropic/Voyage, Gemini, Cohere
LLM Providers	5	OpenAI, Anthropic, Gemini, Ollama, Fallback
Toolbox	24 tools	file, web, data, datetime, text
Rerankers	2	Cohere, Jina

New Document Loaders

Add to src/selectools/rag/loaders.py as new static methods on DocumentLoader. Refactor to loaders/ subpackage with __init__.py re-exporting everything to support SaaS loaders as separate files.

Loader	Method	Dependencies	Complexity	Why it matters
CSV	`from_csv(path, content_columns, metadata_columns)`	stdlib `csv`	Small	Most common structured data format
JSON/JSONL	`from_json(path, text_field)` / `from_jsonl(...)`	stdlib `json`	Small	Standard for API responses, logs, datasets
HTML	`from_html(path_or_content, extract_text=True)`	`beautifulsoup4` (optional)	Small	Web scraping output, saved pages
URL	`from_url(url, timeout=30)`	`requests` + `beautifulsoup4`	Small	Direct URL-to-document (2nd most requested after PDF)
Markdown w/ Frontmatter	`from_markdown(path)`	`pyyaml` (optional)	Small	Static sites, docs, wikis
Google Drive	`from_google_drive(file_id, credentials)`	`google-api-python-client`	Medium	Most-used enterprise doc platform
Notion	`from_notion(page_id, api_key)`	`requests` (existing)	Medium	2nd most-requested SaaS loader
GitHub	`from_github(repo, path, branch, token)`	`requests` (existing)	Small	Developer docs and code
SQL Database	`from_sql(connection_string, query)`	`sqlalchemy` (optional)	Medium	Enterprise data in databases

New Vector Stores

New files in src/selectools/rag/stores/. Each follows the same pattern as chroma.py: inherit VectorStore, implement add_documents, search, delete, clear, lazy-import the dependency. Register in VectorStore.create() factory.

Store	File	Dependencies	Complexity	Why it matters
FAISS	`faiss.py`	`faiss-cpu`	Medium	De facto standard for local high-perf vector search (millions of vectors)
Qdrant	`qdrant.py`	`qdrant-client`	Medium	Fastest-growing vector DB, excellent filtering, cloud + self-hosted
pgvector	`pgvector.py`	`psycopg2-binary`	Medium	Use existing PostgreSQL — no new database needed
Weaviate	`weaviate.py`	`weaviate-client`	Medium	Popular cloud vector DB with GraphQL API
Redis Vector	`redis.py`	`redis` (existing)	Medium	Leverages existing Redis connection from `cache_redis.py`

New Toolbox Modules

New files in src/selectools/toolbox/. Follow @tool decorator pattern, register in get_all_tools() and get_tools_by_category().

Module	Tools	Dependencies	Complexity	Why it matters
`**code_tools.py`	`execute_python`, `execute_shell`	stdlib `subprocess`	Medium	#1 most-used tool in agent frameworks
`search_tools.py`	`google_search`, `duckduckgo_search`	`duckduckgo_search` (optional)	Small-Medium	#2 most-used tool category
`github_tools.py`	`create_issue`, `list_issues`, `create_pr`, `get_file_contents`	`requests` (existing)	Medium	Developer workflow automation
`db_tools.py`	`query_database`, `list_tables`, `describe_table`	`sqlalchemy` (optional)	Medium	Enterprise data access

Dependency Management

All new dependencies are optional and lazy-imported. Add to pyproject.toml:

[project.optional-dependencies]
rag = [
    # existing deps ...
    "beautifulsoup4>=4.12.0",
    "faiss-cpu>=1.7.0",
    "qdrant-client>=1.7.0",
    "psycopg2-binary>=2.9.0",
    "weaviate-client>=4.0.0",
]

Individual stores/loaders remain installable a la carte: pip install selectools faiss-cpu works without the full [rag] group.

Feature	Status	Impact	Effort
Multimodal messages	🟡	High	Medium
OTel observer	🟡	High	Medium
Azure OpenAI provider	🟡	High	Small
Langfuse observer	🟡	High	Small
FAISS Vector Store	🟡	High	Small
Qdrant Vector Store	🟡	Medium	Small
pgvector Store	🟡	High	Small
CSV/JSON/HTML/URL Loaders	🟡	High	Small
Code Execution Tools	🟡	High	Medium
Web Search + GitHub Tools	🟡	High	Small
Database Query Tools	🟡	Medium	Small

Backlog (Unscheduled — Priority Ordered)

Research basis: Competitive analysis of Agno (39k stars), PraisonAI (6.9k stars), and Superagent (6.5k stars) conducted 2026-04-10.

Strategic thesis: selectools wins on depth (50 evals, 7 vector stores, graph orchestration, pattern agents). Close the breadth gap cheaply, own the "production-ready" narrative, adopt the emerging A2A standard.

Reorganized 2026-06-12 against the v0.26.0 codebase: 11 of the 13 original P0–P2 items have shipped. Shipped items moved to the ledger below; the open remainder is re-prioritized for the post-1.0 release train.

Shipped from this backlog

Item	Shipped	Where
Tool-call loop detection (3 detectors, two-tier response)	v0.22.0	`loop_detection.py`, `docs/modules/LOOP_DETECTION.md`
Agentic memory — `remember` tool	v0.16.0	`toolbox/memory_tools.py` (auto-injected with `knowledge_memory`)
Agent-as-API (production REST, auth, SSE)	v0.24.0 (#68)	`serve/api.py`
LiteLLM provider (100+ models)	v0.24.0 (#74)	`providers/litellm_provider.py`
Cost-optimized model router	v0.24.0 (#75)	`providers/router.py`
A2A protocol (server + client + agent cards)	v0.24.0 (#76)	`a2a/`
Toolbox expansion to 48 tools (calculator, email, PDF, Slack, Notion, Linear)	v0.24.0 (#77)	`toolbox/`
Tool result compression	v0.25.0 (#87)	`ToolConfig.compress_results`, `agent/_tool_executor.py`
Session history search (FTS across 4 backends)	v0.24.0 (#79)	`sessions.py` `SessionStore.search()`
Memory tiering / auto-promotion (`UnifiedMemory`, standalone)	v0.24.0 (#78)	`unified_memory.py`
Agent-level HITL / approval	v0.25.0 (#88)	`ToolConfig.require_approval` + `approval_handler`
Planning-as-config	v0.25.0 (#86)	`AgentConfig.planning` (`PlanningConfig`)
Agentic memory — `recall` tool	v0.27.0 (#109)	`toolbox/memory_tools.py` `make_recall_tool` (auto-injected with `remember`)
UnifiedMemory AgentConfig wiring	v0.27.0 (#111)	`MemoryConfig(unified=True, ...)`, `agent/core.py`
Toolbox: Discord, S3, browser, image-gen (48 → 56 tools)	v0.27.0 (#110)	`toolbox/{discord,s3,browser,image}_tools.py`
Cache-rate cost support (OpenAI + Gemini)	v0.27.0 (#112)	`pricing.calculate_cost_with_cached_input`, `cached_prompt_cost`
Cron / scheduled agents	v0.27.0 (#113)	`scheduler.py` (`AgentScheduler`, `cron`, `every`)
Reasoning-as-tool	v0.27.0 (#114)	`toolbox/reasoning_tools.py` (`make_reasoning_tools`, `ReasoningTools`)
Episodic memory retention config	v0.27.0 (#111)	Delivered with the UnifiedMemory wiring — `MemoryConfig.episodic_retention_days` flows through and `add_turn` auto-prunes (tested: `test_retention_pruning`)

Shipped 2026-06-13: the four v1.1 candidates (#109-#112) plus the top two Future/Watch items (cron #113, reasoning tools #114), folded into the v1.0 train. Episodic retention config (#111) needed no separate work — the UnifiedMemory wiring already exposed and auto-applied it. gemini-embedding-2 decision: GA/recommended-for-new; default stays gemini-embedding-001 (incompatible embedding space). See CHANGELOG.md [Unreleased].

Active Queue — needs a product decision before building

The autonomously-buildable, high-conviction items are now shipped. Each remaining item carries a scoping/product call (flagged), so these wait on John's direction rather than getting built blind.

Item	Source	Decision needed	Effort
More DB backends — MongoDB (#116) + DynamoDB shipped; Firestore on demand	Agno	MongoDB + DynamoDB session stores shipped 2026-06-13. Firestore next only if there's demand (needs `google-cloud-firestore`).	Medium each
ML-based guard models	Superagent	Heuristic tier shipped: `PromptInjectionGuardrail` (2026-06-13) covers templated attacks with no model hosting. The model-based tier still needs the hosting decision: bundle a 0.6-4B model, optional extra, or external endpoint?	High
Multi-channel bot gateway	PraisonAI	Roadmap itself says "better as a separate package." In-repo module vs new package = a product/packaging call.	High
Learning system	Agno	Scope is vague (decision logging + preference tracking). Needs a concrete spec before it's buildable.	High
Shadow git checkpoints	PraisonAI	Only relevant if selectools moves toward coding-agent use cases — a direction call.	Medium

Previously Planned (Retained)

Feature	Notes	Target
AWS Bedrock provider	Covered today via LiteLLM; native boto3 wrapper only if enterprise demand	Future
Durable execution / webhooks	Task queue, resume from checkpoint	Future
Code execution sandbox (Docker/E2B)	Sandboxed code execution for untrusted input	Future
Prompt registry / versioning	Version, A/B test, rollback prompts	Future
Time-travel debugging / state replay	Rewind, edit, replay from any checkpoint	v1.x
Voice / real-time audio agents	WebRTC, STT/TTS, sub-500ms latency	v1.x
Rate limiting & quotas	Per-tool and per-user quotas	Future
CRM & business tools	HubSpot, Salesforce integrations	Future
Niche loaders	Confluence, Jira, Discord, Docx	Future
Niche vector stores	Weaviate, Redis Vector, Milvus, OpenSearch, Lance	Future

Previously Planned (Retained)

Feature	Notes	Target
AWS Bedrock provider	Covered today via LiteLLM; native boto3 wrapper only if enterprise demand	Future
Durable execution / webhooks	Task queue, resume from checkpoint	Future
Code execution sandbox (Docker/E2B)	Sandboxed code execution for untrusted input	Future
Prompt registry / versioning	Version, A/B test, rollback prompts	Future
Time-travel debugging / state replay	Rewind, edit, replay from any checkpoint	v1.x
Voice / real-time audio agents	WebRTC, STT/TTS, sub-500ms latency	v1.x
Rate limiting & quotas	Per-tool and per-user quotas	Future
CRM & business tools	HubSpot, Salesforce integrations	Future
Niche loaders	Confluence, Jira, Discord, Docx	Future
Niche vector stores	Weaviate, Redis Vector, Milvus, OpenSearch, Lance	Future