Architecture Patterns
The goal of this page is not to present one "correct" architecture. It is to give customers a set of practical patterns they can adapt.
Pattern 1: Personal agent on a local machine
Shape: one user, one runtime, local context, optional UI.
User
-> Desktop / TUI / CLI
-> Local Codebolt runtime
-> Local tools, files, browser, terminal, MCP
Best for:
- personal productivity
- developer assistance
- private local-first workflows
Why customers like it:
- minimal setup
- strong local context access
- no central infrastructure required
Pattern 2: Headless business worker
Shape: no interactive UI, just a long-running runtime receiving tasks and triggers.
Schedule / Webhook / Internal app
-> Headless Codebolt server
-> Agents and capabilities
-> Business systems, files, APIs, MCP
Best for:
- scheduled work
- background automations
- CI or internal operations
Why customers like it:
- operationally simple
- easy to run under
systemd, Docker, or a job runner - clean separation between runtime and interface
Pattern 3: Channel-based agent gateway
Shape: external channels feed messages into Codebolt, which routes work to agents and sends replies back.
Slack / Telegram / Email / Other channels
-> Channel plugin
-> Routing Gateway
-> Codebolt agent runtime
-> Reply back to the original channel
Best for:
- meeting users where they already work
- shared assistants for teams
- business workflows that begin in chat tools
Why customers like it:
- one runtime can support many channels
- easier rollout than asking everyone to install a new interface
Pattern 4: Monitoring and reaction loop
Shape: external systems emit signals, Codebolt analyzes context, then acts or escalates.
Sentry / Logs / Alerts / Linear / Webhooks
-> Plugin or hook
-> Codebolt runtime
-> Analysis, action, coordination, escalation
Best for:
- reliability workflows
- support triage
- operational monitoring
Why customers like it:
- turns passive monitoring into active response
- can attach reasoning and tool use to every alert
Pattern 5: Embedded web application with agent backend
Shape: another product or internal web app owns the interface while Codebolt provides the agent runtime.
Custom web app / portal
-> Server or plugin SDK integration
-> Codebolt runtime
-> Agents, tasks, memory, tools, orchestration
Best for:
- internal AI products
- domain-specific assistants
- workflow apps with embedded AI operations
Why customers like it:
- Codebolt handles runtime complexity
- the product team keeps full control of UX
Pattern 6: Multi-agent coordinator
Shape: one supervisory layer manages multiple specialized agents and longer work programs.
User / Trigger / Business event
-> Coordinator agent or orchestration flow
-> Specialist agents
-> Shared tasks, jobs, memory, tools, event streams
Best for:
- structured multi-step work
- virtual team models
- longer-running delivery or operations processes
Why customers like it:
- more scalable than one giant general-purpose agent
- easier to reason about responsibilities and escalation paths
Pattern 7: Split control plane and execution plane
Shape: one layer manages agents and routing, another layer executes work in sandboxes or remote runtimes.
Users / Apps / Channels
-> Control plane
-> Codebolt routing and orchestration
-> Remote execution provider / containers / sandboxes
Best for:
- stronger isolation
- tenant separation
- ephemeral workloads
- cloud execution near customer systems
Why customers like it:
- better security boundaries
- easier resource management
- cleaner scaling for heavier workloads
Choosing between the patterns
The best architecture usually depends on four decisions:
- Where should the runtime live: local, server, sandbox, or cloud?
- What should be the visible interface: Codebolt UI, another UI, or no UI?
- What starts the work: human prompts, schedules, events, or all three?
- How autonomous should the system be: assistant, worker, monitor, or coordinator?