Skip to content

Agents

Agents are the engine of Sonora. Each one reads an account or contact, answers a specific question, and writes the result back to the record — health, renewal motion, an escalation risk, a competitor mention, a LinkedIn URL. They run every day across your whole book, so the work surfaces without anyone doing the research by hand.

You use agents two ways: enable one from the built-in library with a click, or write your own. Either way, the output becomes a queryable field that chat, Canvas, and actions all build on.

Each agent is three things: a prompt, a set of data sources, and an output format.

The prompt describes what to find — in natural language, the same way you’d brief an analyst.

Data sources are scoped per agent. Internal sources cover everything Sonora has ingested: support tickets, meeting summaries with sentiment, product usage, feature requests, contact engagement. Web sources cover company sites, news, job postings, funding databases, and press releases. A tech-stack agent might use only web search; a renewal-risk agent might draw entirely from internal usage and support history.

The output format decides what gets written back: a number, a date, a yes/no, a URL, a free-text summary, or a value picked from a list you define. Structured outputs stay queryable from chat and Canvas; free-text outputs are for narrative briefs.

Results attach to customer or contact records. Run an agent once to backfill, or schedule it to refresh on a cadence.

Sonora ships more than thirty pre-built agents. Enable the ones that match how your team works; each is a starting point you can duplicate and adjust. They group into a few families.

Web-research agents that watch the world outside your stack.

AgentDetects
Exec ChangesC-suite and VP-level leadership changes
FundingNew rounds, IPOs, acquisitions
Layoff CheckWorkforce reductions and WARN notices
M&A ActivityMergers or acquisitions involving the account
Pre-Call BriefA combined brief — recent news, health, open issues, next steps — before a meeting
AgentOutput
Renewal PlayClassifies the motion: auto-renew, risk, or upsell
Renewal IntentReads signals for likelihood to renew
Expansion OpportunitiesFlags accounts ripe for upsell
Multi-Product ActivationWhether the account uses three or more key features
Roadmap MatchWhether their open feature requests match what’s planned
AgentOutput
Escalation RiskLikelihood a support issue boils over
Low Engagement WarningAccounts with no recent meetings or support contact
Onboarding LaggardNew accounts stalling before first value
Bad CallAccounts with an eCSAT below 4 in the last two weeks
Sentiment TrendThe direction of customer sentiment over time
AgentOutput
Usage TrendThe direction of product usage
Call Count (90d)Meetings held in the last quarter
Hours SpentTime invested in the account
Last Call DateDate of the most recent meeting
AgentDetects
Competitor MentionsCompetitors named on calls or in tickets
Feature MentionSpecific features coming up in conversations
Unfulfilled Feature RequestsRequests open 90+ days or flagged high priority
G2 Review CandidateHappy contacts worth asking for a review
AgentOutput
Handoff (MEDDPICC)A MEDDPICC brief assembled from the closed deal and early calls
Economic BuyerIdentifies the budget owner on the account

Contact-level agents that run on the people table.

AgentFinds
LinkedIn LookupA contact’s LinkedIn profile URL
Job Title LookupCurrent title
Public MentionRecent news, interviews, or notable posts
Champion StatusWhether a contact reads as a champion or a detractor
Recently Promoted / Recently DepartedJob changes among your contacts

Three agents run automatically on every account, with no setup. Their outputs feed the others above.

eCSAT (estimated CSAT) scores every support interaction 0–10, with the detected emotions and themes that drove the score. eCSAT rolls up into a per-customer trend, so the relationship slope is visible before any human flags it.

Sentiment detects emotional tone in calls and tickets — a label from very negative to very positive, the specific emotions, and the key themes.

Escalation prediction estimates the probability that a support issue escalates, based on tone, issue age, and repeat contacts.

Customer health is the agent most teams configure first. It distills the signals you choose into a single 1–10 number per account. Configure each signal independently — pick from product usage, meeting cadence, support volume, sentiment, renewal intent, or a custom enrichment from a custom agent. Each signal scores 1 (critical) to 10 (excellent). Each gets a weight from 0 to 10.

The final score is a straight weighted average:

Health Score = Signal₁ · Weight₁  +  Signal₂ · Weight₂  +  …Weight₁  +  Weight₂  +  …

Worked example: Product Usage scores 8 at weight 10, Touchpoints scores 6 at weight 5, Sentiment scores 7 at weight 5. That gives (80 + 30 + 35) ÷ 20 = 7.3.

Signals with no available data drop out of both numerator and denominator, so missing inputs don’t drag the score down.

The 1–10 score maps to a status through configurable bands. The defaults:

StatusScore range
Thriving8 – 10
Healthy6.5 – 8
Emerging Concerns4 – 6.5
At Risk0 – 4

Drag the threshold sliders in settings to match how your team actually talks about accounts. Every customer reclassifies instantly.

Some signals matter so much that a single bad reading should override the average. Toggle Critical Override on a signal and set a threshold; if that signal scores at or below the threshold, the account is flagged At Risk regardless of the weighted score. A customer with strong usage and sentiment but an explicit non-renewal signal still surfaces — which is the point.

When nothing in the library fits, write your own: a prompt, the data sources it can read, an output format, and when it runs. Common uses:

  • Tech stack and competitor signals — what a customer uses, what they replaced, who else they’re evaluating
  • ICP scoring — a fit score from public company data plus your account criteria
  • Renewal context — usage trajectory, support load, sentiment change, last meaningful touchpoint
  • Custom qualification — any rubric your team scores accounts against by hand today

Pick a structured output (number, date, boolean, list selection) when you want the result in dashboards and chat queries, or free text for a narrative brief.

Agent results feed the rest of the product:

  • Actions trigger on agent values. When health drops to At Risk or escalation probability crosses your threshold, a playbook fires and an action appears with context and a suggested next step.
  • Chat treats agent outputs as queryable fields. “Show at-risk customers” pulls health; “show declining sentiment” pulls eCSAT trends.
  • Canvas blocks render them — health distributions, satisfaction trends, the output of any custom agent.
  • The customer profile shows them inline, with the date each was last computed and the data the agent used. Health and CSAT scores sync back to Salesforce or HubSpot so your team sees them where they already work.

Agents that touch the web cost more than ones that read internal data only. Sonora rate-limits per-tenant runs to keep costs predictable; if you need a one-off backfill across thousands of records, talk to your account team first so we can scale up the budget for the run. See credits and rate limits for how this work is metered against your plan.