How It Works

From SOP to running agent in one afternoon.

Nexwatt's three-phase lifecycle takes your existing process documentation — a Notion page, a SOP, a Loom transcript — and turns it into a live, monitored agent. No YAML. No trigger chain maintenance. No engineering ticket to change a routing rule.

Try It Free 5-Minute Quickstart

Phase 1

Describe. Write or paste your process in plain English.

Give Nexwatt what you already have — a Notion page, a Google Doc, a Loom recording transcript, or a bullet list you wrote in Slack. You don't need to format it for the system. The parser reads it as prose.

Nexwatt extracts the steps, identifies the tools being referenced by name, detects conditional phrases ("if the deal is over $50k, route to the enterprise CS lead"), and flags anything it's uncertain about for you to clarify.

What the parser extracts

Sequential steps and their ordering
Tool names and referenced systems
Conditional logic and branching points
Exception cases and escalation triggers
Human roles referenced ("the manager", "CS lead")

nexwatt ingest onboarding-sop.txt

Parsing... done

Steps extracted: 7
Conditions: 3
Tools: BambooHR, Jira, Slack, Google Workspace
Ambiguous: 1 (step 4 — "notify the right person" — clarify)

Review extracted logic? (y/n): y

Step 1: Trigger — New hire record in BambooHR
Step 2: Action — Create Jira onboarding epic
Step 3: Condition — If dept = Engineering, add dev env task
Step 4: [CLARIFY] Who is "the right person"?

nexwatt simulate onboard-agent-v1

Generating test cases...
Test cases created: 40

Running simulations...

Pass: 37 / 40
Low confidence: 3 runs flagged

Flagged paths:
— Step 4: dept=Legal → no matching CS rule (52%)
— Step 6: Jira API timeout scenario (58%)
— Step 7: Manager field empty in BambooHR (49%)

Threshold set to 60%. 3 paths will escalate.
Ready for deployment review.

Phase 2

Build and Test. Simulations before anything touches live data.

Nexwatt generates 40+ synthetic test cases from your process doc and runs the agent against them in simulation mode. No live API calls, no data changed — just the agent logic running against representative inputs.

Paths that produce confidence below your threshold are flagged with the exact reason: missing data, ambiguous condition, or API unavailability. You review each flagged path and either clarify the rule or accept that it escalates to a human.

Live runs

Real-time execution

Agent listens for triggers, runs each step sequentially, logs every action, and routes exceptions to the review queue automatically. No polling required.

Escalation queue

Humans see the edge cases

Low-confidence items go to a shared review queue. Team members see the context, make a decision, and the agent calibrates based on that resolution — improving over time.

Audit trail

Every decision logged

Timestamp, confidence score, input data (redacted per your policy), action taken, outcome. Available in the dashboard or via API export to your data warehouse.

Common Questions

Questions before the first deploy.

What if my process changes after the agent is deployed?

You re-ingest your updated process doc. Nexwatt diffs the new version against the deployed agent, highlights what changed, and asks you to confirm before updating. The old agent version is preserved in history. You can roll back at any time.

Can agents chain together — one agent triggering another?

Yes. Agents can emit events that trigger other agents. You can define a "deal closed" agent that triggers an "onboarding prep" agent, for example. The full chain is visible in the audit log with a shared run ID across both agents.

What happens when the agent makes the wrong call?

Every automated action is logged with enough context to reverse it manually. Most tool connectors support undo operations via the API (delete ticket, unsend notification where available). You flag the run as incorrect from the audit log, and the agent treats it as a calibration signal.

How do I connect my tools?

Pre-built connectors require OAuth authorization — click to connect, grant permissions, done. For tools not in our catalog, you define a custom REST connector by providing the base URL, auth method, and the endpoints the agent should use. Most engineering teams can do this in under an hour.

Is my process data stored or used to train models?

Your process data and the data your agents process are not used to train shared models. Calibration data is scoped to your account. Input data in audit logs is redacted at the field level per your data handling policy. See our Security page for the full data architecture.

What languages and regions are supported?

Process ingestion works best with English-language documentation. Early access is available to US-based companies only. We're working on international availability and will announce when it's ready.

Start with one process. See it running today.

Get Early Access 5-Minute Quickstart Guide