An agent that knows when to act and when to ask.
Nexwatt's orchestration engine reads your process documentation, builds the agent logic, connects to your tools via pre-built and custom REST connectors, and runs with per-step confidence scoring built in. Every decision is logged. Every exception routes to a human queue — not an inbox.
Describe your process in plain English. That's the whole input.
Nexwatt reads your existing documentation — SOPs, Loom transcripts, Notion pages, even a bullet list in an email. You don't need to rewrite anything for the system. The parser extracts steps, conditions, branching logic, and tool references automatically.
Accepted formats: plain text, PDF, DOCX, Markdown, or a public URL. Nexwatt identifies tool names, conditional phrases ("if the deal is over $50k…"), and exception cases within the existing prose.
Parsing document...
File: deal-closed-sop.txt (1.4 KB)
Extracted steps:
1. Trigger: Deal status changes to Closed Won
2. Lookup: Check Jira for open implementation tickets
3. Condition: If tickets > 0, assign to CS lead
4. Create: Slack notification #deals-closed
5. Update: Google Sheets deal tracker row
Tools detected: Salesforce, Jira, Slack, Google Sheets
Conditions found: 2
Exception paths: 1 (no CS lead available)
Ready to build agent. Review extracted logic? (y/n)
Review the decision tree before anything goes live.
Step-by-step canvas view
See every step in the agent's decision tree laid out sequentially. Each step shows the tool it calls, the condition it evaluates, and the action it takes. Edit any step before running simulations.
Simulation mode
Run the agent against synthetic test cases before anything touches live data. Nexwatt generates 40+ scenarios from your process doc and flags every step that produces a low-confidence result or ambiguous outcome — with the exact reason: missing data, ambiguous condition, or API unavailability.
Per-step confidence thresholds
Set the confidence floor independently for each step. A deal-routing step might run at 70%; a payment-touching step might stay at 90%. Conservative ops teams start high and lower thresholds as the agent proves out. You control the risk tolerance — not a global switch.
Conditional branches and retry logic
Define what happens when a condition isn't met, when an API returns an error, or when a required field is empty. Fallback chains can route to a different tool, wait and retry, or escalate immediately.
Confidence scoring means the agent knows what it doesn't know.
Every agent decision carries a confidence score derived from the clarity of the input data, the specificity of the matching rule, and the outcome history of similar prior decisions.
When confidence drops below the configured threshold, the item routes to the human escalation queue — not your Slack DMs. You review, decide, and the agent updates its calibration based on that resolution.
- 01Rule match scoring — how precisely the input matches a defined condition
- 02Data completeness — missing fields reduce confidence before the step runs
- 03Historical calibration — prior human resolutions inform future scoring
Every decision logged. No black boxes.
Your compliance team, your InfoSec reviewer, and your ops lead all read the same log — timestamp, step, confidence score, action, outcome. Input data is redacted at the field level per your data policy. Export to CSV or push to your data warehouse via the API.
Input data redacted per policy. Full payload available to authorized roles.
Deploy your first agent this week.
Pilot tier is free — one agent, 500 runs/month, three integrations. Or read the docs first to see exactly what the orchestration engine can handle.