Built different,
runs in parallel.
Mode Agent isn't a chat box. It's a full orchestration layer that conducts the models and tools you already use, with specialized agents working alongside you in parallel. This is the machinery of a self-driving company. Here's what powers it.
Parallel agent execution
Every goal is broken into specialized roles and launched simultaneously. Your researcher doesn't block your marketer. Your marketer doesn't block your build. Work that used to happen sequentially now happens in parallel, the way a real team works.
- Configurable agent roster per project type
- Shared signal bus for inter-agent communication
- Each agent gets scoped tools, model, and permissions
- Results merged and diff-ready for your review
Cross-Provider Routing
Mode Agent auto-routes each agent to the optimal model based on task type, cost, and latency. The best model for every task, automatically. It gets better every time any model does. Swap providers with a one-line config change.
- Intelligent auto-routing assigns the best model per agent role
- 700+ models accessible
- Per-agent model assignment in
.mode/config.json - Hot-swap models without restarting agents
tmux Integration
AI sessions that persist and travel with you. Start a run on your laptop, detach, and resume from any device. Phone, remote server, a colleague's machine. Your agents keep running while you're gone, and you pick up exactly where you left off.
- Sessions persist after disconnect, and agents keep running
- Reattach from any device with SSH access
- Full session history scrollback on resume
- Works with your existing tmux config and theme
Sandboxed execution
Every agent runs inside an isolated environment where it physically can't touch anything outside its scope. You define the boundaries once, then agents operate freely within them. No permission prompts interrupting your flow. The sandbox itself is the safety layer.
- Isolated containers per agent with scoped file system access
- Define allowed commands, files, and network access per sandbox
- Agents run freely inside their boundaries with zero prompts
- All activity logged automatically for full auditability
# config for coder.1 (limited scope) "coder:1": { "fs": { "roots": ["./src"] }, "bash": { "allow": ["npm", "git"] } } # agent tries to escape scope [coder:1] Attempting rm -rf / ✗ BLOCKED. 'rm' not in allow-list ✗ LOGGED. Attempt saved to audit log [coder:1] Attempting write ../../../.env ✗ BLOCKED. Path outside fs.roots
Work Graphs
Mode Agent decomposes your goals into dependency-aware task graphs, not flat to-do lists. It identifies which tasks can run in parallel and which must wait for dependencies, then dynamically adjusts the graph as work progresses and new information emerges.
- Automatic dependency detection between tasks
- Maximum parallelism with correct ordering guaranteed
- Dynamic re-planning when tasks finish early or produce unexpected results
- Visual task graph in your terminal showing progress and dependencies
LLM Council
Critical decisions shouldn't depend on a single model's opinion. LLM Council convenes multiple models to deliberate, challenge, and validate each other. Multiple models cross-check each other's reasoning. Blind spots one model misses, another catches.
- Multiple models independently evaluate the same question
- Cross-model validation catches hallucinations and blind spots
- Configurable quorum rules including majority vote, unanimous, or weighted
- Confidence scores based on inter-model agreement
Context engine management
Every model has a different context window, from 16K to 128K to 1M to 2M tokens. Mode Agent's Context Engine intelligently manages what each agent sees by chunking large codebases, compressing history, and prioritizing the most relevant context for each task.
- Automatic context sizing per model's window capacity
- Intelligent chunking that splits codebases by dependency graph
- Rolling compression for long-running agent sessions
- Priority-ranked context, from recent edits to related files to full repo
Continuous autonomous operation
Agents don't stop when you close the terminal. Mode Agent runs continuously, working through tasks autonomously, picking up the next item, running tests, and iterating without waiting for a new prompt each time.
- Agents work through task queues without prompting
- Automatic iteration on test failures and review feedback
- Background execution survives disconnects
- Progress notifications when milestones are hit
Cloud execution
Most AI coding tools die when you close the terminal or put your laptop to sleep. Mode Agent doesn't. When you start a run, it executes in the cloud. Close your laptop, switch devices, go to bed. Your agents keep working. Come back and everything is done, with a full log of every decision made while you were away.
- Agents run on cloud infrastructure, not your machine
- Survive disconnects, sleep mode, and network drops
- Resume and monitor from any device
- Full execution log of everything that happened while you were away
Universal API
Coming soon with separate per-call pricing
Developers hit one endpoint with a goal. Mode Agent decomposes it into tasks, selects the right model for each agent, assigns agents to every subtask, and runs them all in parallel. The API learns which models work best for which tasks over time, getting smarter with every call. Embed the Agent's intelligence into any product and stream results back in real time.
- Single endpoint for any task
- Intelligent model selection per agent
- Real-time streaming results
- Embed in any product or workflow
Custom agents
Create specialized agents tailored to your business. Define the system prompt, allowed tools, model preference, and scope for each one. Build a brand voice agent that enforces your tone across every piece of content. Build a compliance checker that reviews every PR against your regulatory requirements. Custom agents run alongside built-in agents in the same parallel workflows, giving you total control over how Mode Agent works for your team.
- Define agent roles with custom system prompts
- Assign specific models and tools per agent
- Share agents across your team
- Run custom agents alongside built-in ones
# mode.config.json { "agents": { "brand-voice": { "prompt": "Enforce our brand tone guide...", "model": "anthropic:claude-sonnet-4-6", "tools": ["read", "write", "lint"] }, "compliance": { "prompt": "Review against SOC2 controls...", "model": "anthropic:claude-opus-4-6", "tools": ["read", "search"] } } } $ Review the new blog post Loading custom agents from config... [brand-voice] ✓ Tone aligned, 2 edits suggested [compliance] ✓ No regulatory flags found [editor] ✓ Grammar and clarity pass ✓ Done. All checks passed
Integration hub
Mode Agent connects directly to the tools you already run your business on. Agents don't just generate text and leave you to copy-paste. They push code to GitHub, deploy to Vercel, create issues in Linear, send updates to Slack, and process payments through Stripe. Every integration is authenticated once and available to all your agents automatically.
- Native connections to Stripe, GitHub, Vercel, AWS, and more
- Authenticate once, every agent inherits access
- Agents take real actions, not just suggestions
- Add custom integrations via webhooks and API connectors