Orchestration Landscape

Agent sessions — message routing, memory scoping, topic forking, multi-agent coordination. Routes messages to folders, isolates context per team, composes tools via MCP.

Multi-tenant agent context orchestration: topic-suffix routing (target=corp/eng#alerts), folder-per-team isolation (container boundaries), sibling observation (context sharing without secret leakage), session forking (cp -r .claude/).

• Deploying Claude agents across multiple teams with isolated memory and secrets
• Routing messages from Slack/Telegram/Discord to specific agent folders
• Multi-agent coordination with shared context but scoped permissions
• Self-hosted requirement — no metadata leaves your infrastructure

• Workflow execution guarantees → use Temporal
• Data pipeline orchestration → use Dagster
• App-to-app integration automation → use n8n
• Managed SaaS (arizuko is self-hosted only)

Multi-agent teams with a spatial canvas UI. Visual workspace where agents have status, communicate directly, and coordinate across repos. Desktop-based agent coordination.

Spatial canvas for agent visualization. See your whole team at a glance — who's active, waiting, idle. Agent-to-agent communication (DMs, group channels, structured handoffs) with real-time visual feedback.

• Local development with multiple AI employees across engineering, ops, research
• Visual debugging of multi-agent workflows
• Coordination across repositories (each agent gets its own clone)
• Real-time status monitoring without building dashboards

• Production multi-tenant deployment → use arizuko
• Message routing from external platforms (Slack, Telegram) → use arizuko
• Team-based isolation with container boundaries → use arizuko

Distributed workflows with durable execution. Guarantees workflows complete despite infrastructure failures — persists state at each step, resumes from last checkpoint on a different worker.

Durable execution with replay guarantees. Replaces manual retry logic + state DB + cron scheduler. Event sourcing for crash recovery, structured retry policies, automatic timeouts.

• Payment processing, saga patterns, multi-step transactions
• Microservice coordination with guaranteed completion
• Long-running workflows (hours or days) that must survive restarts
• Worker versioning required

• Simple cron jobs without complex state → use systemd timers
• Real-time event processing at millisecond latency → use Kafka
• Data pipeline orchestration with asset lineage → use Dagster

Data pipelines — tables, datasets, ML models, reports. Asset-first orchestration where data assets are first-class citizens with clear lineage and dependencies.

Asset-first orchestration (not task-first). Unified control plane for data platforms with integrated observability. Multi-tenancy via code locations (subprocess + virtualenv isolation).

• Data engineering teams building data platforms in code
• Asset lineage tracking (which tables depend on which models)
• Multi-tenant data pipeline deployment
• Integration with dbt, Spark, Pandas, Polars

• Python-first dynamic workflows → use Prefect
• Battle-tested batch ETL at massive scale → use Airflow
• General distributed execution guarantees → use Temporal

Python workflows with flexible execution. Task-based model (not asset-first). Runs your Python functions as-is — add @flow decorator and done.

Python-first orchestration with zero marginal cost (charges by users/workspaces, not task executions). Dynamic, event-driven workflows. Operational resilience without an opinionated data model.

• Python-first teams who want flexibility over opinionated best practices
• Dynamic workflows where structure changes at runtime
• Cost-sensitive deployments (no per-execution charges)
• Event-driven automation beyond just scheduled runs

• Asset lineage and data-first orchestration → use Dagster
• Durable execution guarantees → use Temporal
• Legacy Airflow replacement at enterprise scale → use Airflow or Astronomer

Batch data processing — ETL/ELT, data integration, ML training workflows. Battle-tested at scale with well-defined start and end. DAG-based orchestration.

Battle-tested batch orchestration at massive scale. Largest ecosystem for data integrations. Regulatory compliance in financial services (Dodd-Frank, GDPR reporting).

• ETL/ELT at enterprise scale (hundreds of pipelines)
• Financial services regulatory reporting
• ML model training with SageMaker integration
• Existing Airflow investment or expertise

• Real-time streaming (sub-second latency) → use Kafka
• Thousands of tiny tasks every few seconds
• Modern data platform with asset lineage → use Dagster
• Python-first flexibility → use Prefect

Integration workflows — app-to-app automation, API calls, webhooks, form submissions. Node-based canvas with 70+ AI-specific nodes (LLMs, embeddings, vector DBs, OCR).

Self-hosted AI-native automation platform. Charges per workflow (not per task) — dramatically cheaper for complex multi-step automations. LangChain integration with 70 AI nodes.

• Self-hosted requirement — no data leaves your infrastructure
• AI-native workflows (LLM chains, embeddings, vector search)
• Complex automations with loops, branching, parallel execution
• Technical teams comfortable with node-based canvas + JSON

• Non-technical teams → use Zapier
• Need 7,000+ integrations out-of-box → use Zapier
• UI-triggered automation for internal tools → use Retool Workflows

Personal productivity automation — 7,000+ integrations, no-code. Connects applications at the surface level. Charges per task.

Largest integration breadth (7,000+ apps). No-code accessibility for non-technical teams. AI Agents (autonomous AI teammates) and Copilot (AI-powered builder) launched 2026.

• Non-technical teams (marketing, sales, ops)
• Simple 2–3 step automations between common apps
• Time-to-automation more important than cost at scale

• Enterprise governance with audit trails → use Workato
• Self-hosted requirement → use n8n
• Complex automations with loops/branching → use n8n
• Cost-sensitive at scale (charges per task) → use n8n or Prefect

UI-triggered automation for internal tools. Extension of Retool's internal tool builder. Background tasks tied to user interfaces, not a standalone automation engine.

Tight integration between UI components and workflow engine — natural extension for teams already on Retool for admin panels, ops dashboards, approval flows.

• Already invested in the Retool ecosystem
• Automation surfaces inside internal apps (approval button triggers workflow)
• Consolidated stack preferred (one platform for UI + automation)

• Standalone automation platform → use n8n or Prefect
• AI-native workflows → use n8n
• Not on Retool → limited value outside the ecosystem

Agent memory architecture — tiered memory (core/archival/recall). OS-inspired model: core memory (always in-context), archival memory (vector store queried explicitly), recall memory (conversation history).

Tiered memory for stateful agents that learn and self-improve over time. #1 model-agnostic OSS harness on TerminalBench. Conversations API for shared memory across parallel experiences (Jan 2026).

• Long-running agents that persist across sessions
• Agents that continuously learn from experience
• Memory-first coding agents (Letta Code)
• Self-hosted memory (UC Berkeley open-source)

• Short-lived agents with no learning requirement
• Cross-tool organizational memory → use Memory Store
• Agent session orchestration (routing/isolation) → use arizuko

Organizational memory aggregation — cross-tool context syncing (Fathom meetings, Slack, Granola notes, Claude/Codex outputs). Background organization by project/person/decision. MCP-native distribution.

Shared context layer (not internal memory architecture). Multi-source capture, background organization, shared recall across agents and humans. Memory portable across model/client boundaries via MCP.

• Team organizational memory (1–12 users)
• Cross-tool context aggregation (record in Claude, recall in Codex)
• Fragmented documentation problem with no single source of truth
• Living playbooks accessible in AI tools

• Self-hosted requirement (Memory Store is a YC SaaS product)
• Single-agent focus → use Letta for agent-internal memory
• Local-only memory → use Letta or your agent framework's native memory