Skip to main content
Stratus is purpose-built for multi-step tasks with predictable state transitions — situations where an agent needs to plan a sequence of actions, predict what each action will produce, and recover when reality diverges from the plan.

​
Should You Use Stratus?

Use StratusUse LLM directly
Steps3+ sequential actionsSingle-step
EnvironmentPredictable state transitionsChaotic / stochastic
GoalClear, definable end stateOpen-ended generation
RecoveryNeeds error handling & replanningFire-and-forget
Hybrid pattern: Plan with Stratus → generate with LLM. Example: plan a research strategy (Stratus), write the final report (LLM).

​
Use Cases

Task: Book a flight from SFO to NYC for December 15–18, budget under $500.Stratus plans the complete interaction sequence upfront — select origin, select destination, open calendar, pick dates, filter by price — before a single click happens. When a date picker throws a validation error, it replans rather than halting.
Success Rate
Traditional agent~12%
Stratus agent~35%
const response = await client.chat.completions.create({
  model: 'stratus-x1ac-small-gpt-4o',
  messages: [
    {
      role: 'system',
      content: 'Current page: Google Flights. Date pickers and destination fields visible.'
    },
    {
      role: 'user',
      content: 'Book flight SFO to NYC, Dec 15-18, under $500'
    }
  ]
});
// X1 plans: select_origin → select_dest → open_calendar → pick_dates → filter_price
Task: Extract product prices, reviews, and availability across paginated e-commerce listings.Stratus tracks extraction state across pages, handles dynamic content loading, and navigates pagination without losing context — producing consistent structured output even when layouts change.
Result
Speed vs. traditional scraping5–10× faster
Extraction accuracy90%+
Task: “What MCU movies were released in 2019 and who directed them?”Rather than issuing searches reactively, Stratus pre-plans the query chain: find MCU 2019 titles → extract names → look up each director. Intermediate results are tracked, so a failed sub-query triggers targeted recovery rather than a full restart.
HotpotQA Score
Traditional agent~18%
Stratus agent~42%
const response = await client.chat.completions.create({
  model: 'stratus-x1ac-base-claude-sonnet-4-5',
  messages: [
    {
      role: 'system',
      content: 'You have access to Wikipedia search and can make multiple queries.'
    },
    {
      role: 'user',
      content: 'What MCU movies were released in 2019 and who directed them?'
    }
  ]
});
// X1 plans: search_mcu_2019 → extract_titles → search_each_director
Task: Build a knowledge graph of relationships from a research paper.Stratus plans extraction order (authors → institutions → citations → concepts), predicts entity boundaries, and maintains graph state across extractions — preventing duplicate nodes and missed edges.
Task: Process new customer signup: Create account → Send welcome email → Add to CRM → Notify team.Stratus predicts the API response state at each step before calling it. If createUser fails, the downstream steps are skipped before any erroneous calls are made — not after.
const response = await client.chat.completions.create({
  model: 'stratus-x1ac-base-gpt-4o',
  messages: [
    {
      role: 'system',
      content: 'Available APIs: createUser, sendEmail, addToCRM, notifySlack'
    },
    {
      role: 'user',
      content: 'Process signup for john@example.com'
    }
  ]
});
// X1 plans: createUser → sendEmail → addToCRM → notifySlack
// Predicted states: user_created → email_sent → crm_updated → team_notified
Task: Test checkout flow end-to-end across multiple scenarios.Stratus plans test case sequences, predicts expected states at each step, and flags mismatches between predicted and actual state — catching state-dependent bugs that static test scripts miss.
Result
Execution speed3× faster
CoverageFollows all branching paths
Bug detectionCatches state-dependent regressions
Task: Find where authentication logic is implemented, trace its usage, and reproduce a multi-step bug.Stratus plans the search strategy (find auth files → trace imports → find call sites), maintains context across codebase exploration, and predicts system state at each reproduction step — isolating where actual state diverges from expected.

​
When NOT to Use Stratus

Single-step tasks

Simple classification, formatting, or one-shot generation. The world model adds overhead with no benefit.

Open-ended generation

Creative writing, brainstorming, summarization. No clear goal state means prediction has nothing to anchor to.

Highly stochastic environments

Unpredictable or chaotic systems where state transitions can’t be modeled.

Sub-100ms latency requirements

The X1 planning layer adds latency. If real-time reaction is the constraint, use the LLM directly.

​
Next Steps

Quickstart

Make your first X1 call in under 60 seconds.

Architecture

How the X1 world model actually works.

API Reference

Full endpoint docs and parameters.