Lesson · 6 min read
Your first thread
A thread is a conversation with your agent. This lesson covers how to open one well, what the agent needs to act autonomously, and how to structure requests that produce verifiable output.
What a thread is
A thread is a bounded context window: the agent reads everything in it, reasons over it, acts on it, and writes its receipts back into it. Each thread is independent. The agent carries its configured skills, memories, and approval gates into every thread — but one thread's chat history does not bleed into the next.
You can open a thread manually from the portal, trigger one via a scheduled cron, fire one from a webhook, or dictate one by voice. The interaction model is the same regardless of trigger.
What the agent needs to act
Strong opening messages share three things:
- 01
The outcome, not the steps
Tell the agent what done looks like, not how to get there. "Pull the A/R aging report from QBO and flag anything over 60 days" is stronger than "go to QuickBooks, click Reports, click A/R Aging." The agent routes to the right skill and decides the steps.
- 02
The approval gates
If the action has consequences you want to review first, say so. "Draft the collection email but hold for my approval before sending" invokes the approval gate. Without that phrase, the agent will execute autonomously if the action is within its configured autonomous threshold.
- 03
The verification criterion
What does correct output look like? "Verify by pulling the QBO JE back and reading the line items to me" or "Confirm with a screenshot of the sent folder" tells the agent how to prove it's done. Agents that do not verify are reporting what they intended, not what happened.
Example opening messages
Three patterns that work across different skill domains.
Pull the QBO A/R aging for Prospectr. Flag anything over 60 days. Draft a collection email for each flagged account in my voice, hold each draft for approval before sending. Verify by pulling the draft list back and reading me the subject lines.
Check the inbox for any new HOT lead replies since 6 AM. For each one: score it, enrich the phone number, and add it to the delivery queue for my review. If the reply is an opt-out, suppress it immediately without flagging me. Report back with a count.
Run the weekly deliverability check on all active Instantly campaigns. Flag any with send-to-lead ratio above 5000:1 or bounce rate above 5%. Draft a recommended copy change for each flagged campaign, do not push anything live. Show me the list.
Using approval gates inline
Your agent ships with configurable approval thresholds in its agent configuration. Those thresholds are the floor — any action above them is gated by default. You can also gate individual actions inline with language in the thread:
- ▸ "Draft and hold for my approval"
- ▸ "Stage this but do not send"
- ▸ "Queue it, I will review before 5 PM"
- ▸ "Do not post — email me the preview first"
Any approval request is time-bounded. If you do not respond within the configured window, the action stays held — it does not auto-release.
Verifying output
Every action your agent takes produces a receipt in the thread: what ran, what changed, what it cost. Receipts are not proof — they are the agent's account of what it intended to do. To verify actual output, ask the agent to pull from the system of record:
- ▸ QBO Journal Entry: re-pull by ID and read back line items, accounts, and amounts
- ▸ Email draft: fetch from Gmail Drafts, confirm recipient, subject, and body match intent
- ▸ CRM write: re-fetch the contact and confirm the field update rendered correctly
- ▸ Bulk data operation: sample first, middle, last, and two random records