Skip to main content
PER-DOMAIN AGENTS

Instead of one giant assistant that knows everything, ParsaLink runs a fleet of small, focused agents. Each one owns a single CRM domain — and the right one is picked for you based on what you ask.

​
Why Multiple Agents?

A single all-knowing assistant has to load every tool, every example, and every instruction every time it answers — even when you’re just asking it to mark a task complete. That’s expensive (lots of LLM tokens), slow, and less accurate. ParsaLink routes each request through a small classifier first, identifies the right domain (contacts, opportunities, email, etc.), and then loads only that domain’s tools and system prompt. You get faster, cheaper, more accurate responses. The agents share a common base prompt (formatting rules, response style, when to confirm vs. act) but each one has domain-specific knowledge layered on top.

​
The 16 Domains

AgentWhat it owns
contactsSearch, create, update, bulk import, merge, duplicate checks for Contact records
accountsSame shape, for Account (company) records
opportunitiesDeal CRUD, stage moves, probability + amount updates, related contact/account lookup
casesSupport case CRUD, priority + status transitions
tasksTask CRUD + reminders, due-date logic, automation-spawned tasks
emailDrafting, sending, template creation/cloning, email-account routing
automationsBuilding new automation rules from a natural-language description
meeting-notesReading and writing meeting note attached to calendar events
tagsTag CRUD across all entity types, segmentation queries
commentsThreaded comments on contacts, accounts, opportunities, cases, tasks
analyticsRunning ad-hoc analytics queries (generates safe Django ORM Python under the hood)
pipelinesPipeline + stage CRUD, opportunity pipeline configuration
communicationsTwilio call logs + SMS context lookup
calendarBooking events, comparing team availability, sending invites
webWeb search for facts outside the CRM
shared-baseCommon ground rules every other agent inherits
The classifier may activate multiple namespaces for one message. Asking “Show me all the qualified contacts at Globex with an open opportunity, and draft them an intro email” loads contacts, accounts, opportunities, and email together.

​
How the Router Decides

The router classifier (router-classifier prompt template) takes your raw message and returns a JSON object like: 
{
  "intent": "send_intro_email",
  "namespaces": ["contacts", "email"],
  "confidence": 0.96,
  "reason": "User wants to email new contacts in a segment"
}
The picked namespaces’ tools and system prompts are then loaded for the actual answer.
You’ll usually see this happen invisibly. In some borderline cases the agent may ask a quick clarifying question if the classifier is unsure (e.g. “Did you mean to update the contact or the opportunity?”).

​
Customizing the Agents (Admin)

All 16 prompts live in Admin → LLM → Prompt Templates with template_type = ai_agent. They’re versioned and audit-logged, so tuning the wording of a domain agent doesn’t require a code deploy. Pick a template, edit the body, save. The change applies on the next message — no restart needed.
Editing the agent-shared-base template affects every domain agent. Test changes on a non-production workspace first if you’re making big structural edits.

​
Adding Capabilities

Each agent gets its set of tools from the matching directory under packages/django/crm/ai_tools/ — contacts.py, accounts.py, etc. Adding a new capability to an agent is a code change (new function decorated with the tool registration). New tools are surfaced to the LLM automatically once the agent’s namespace is loaded. If you have an idea for a capability that doesn’t fit a domain — open a request and we’ll figure out where it belongs.

​
Reading the LLM Requests

Every agent call lands in Admin → LLM → Requests with: which template was used, which model, the input message, the response, token counts, cost. Useful for:
  • Debugging when an agent answers wrong (“it should have used update_task but used create_task”)
  • Spot-checking which namespaces were loaded for a borderline message
  • Measuring spend per domain (e.g. “the email agent is our biggest spender, time to tighten its prompt”)

​
What Each Agent Can Actually Do

The table above is the elevator pitch. Below is what each agent can concretely accomplish when you ask it something — useful for picking which messages to send into chat (vs. doing it manually).

​
contacts

  • “Find John from Acme” — search by partial name + company; disambiguation prompts when multiple match
  • “Show me everyone tagged VIP who hasn’t been emailed in 30 days” — combined filter on tags + last activity
  • “Create a contact for Sarah Smith, VP Marketing at Stitch, [email protected]” — single-line create
  • “Merge the two Globex Sarahs and keep the most recent email” — duplicate cleanup
  • “Bulk-update everyone in the prospect stage to source=Cold Outreach” — mass field updates

​
accounts

  • “Find all accounts in the SaaS industry with >100 employees” — combined firmographic filter
  • “What’s the latest activity on the Globex account?” — timeline summary across linked contacts and opportunities
  • “Show me accounts with no opportunity in the last 90 days” — stale account triage

​
opportunities

  • “Show me deals closing this quarter” — closing-window filter
  • “Move the Globex deal to Negotiation and set probability to 70%” — stage + probability update
  • “What’s the total amount of open opportunities by stage?” — pipeline summary

​
cases

  • “Show me urgent cases that are still open” — priority + status filter
  • “Escalate this case to Andrew and tag it vip-customer” — assign + tag in one step

​
tasks

  • “What’s overdue for me?” — personal overdue triage
  • “Push everything on Amy’s plate due Friday to next Monday” — bulk reassign + reschedule
  • “Create a task to follow up with the Globex contacts in 3 days” — relative-date scheduling

​
email

  • “Draft an intro email to Sarah at Stitch about our marketing audit” — AI compose grounded in workspace knowledge
  • “Send the proposal email template to John Smith with the variable = $12,500” — template variable injection
  • “Clone the Proposal Follow-Up template, rewrite it for cold outreach, save as v2” — template authoring without leaving chat
  • “Show me which template has the best open rate this month” — per-template analytics

​
automations (build via the automation-builder template)

  • “Build a rule that sends a welcome email when a contact’s source is set to Website, then creates a follow-up task in 3 days” — describe the workflow in plain English, the agent assembles the trigger + steps and saves the rule
  • “Show me which automation rule sent the most emails last week” — analytics on rule firing

​
calendar

  • “Find a 45-minute slot for Andrew, Amy, and the Globex contacts next week, afternoons” — multi-user availability intersection
  • “Reschedule today’s 3pm to tomorrow at 11am, keep the same attendees” — single-line reschedule
  • “Book a Teams meeting with John Smith tomorrow at 10am and send him the invite” — provider-agnostic booking (Google Meet or Teams)

​
meeting-notes

  • “Save these notes on the Globex meeting and create a task to send the case study” — paste raw notes, agent links them to the right event + extracts action items
  • “Summarize what we discussed with Stitch last week” — surfaces meeting notes filtered by account

​
tags

  • “Show me everyone tagged enterprise AND interested” — combined tag filter
  • “Create a renewal-2027 tag and apply it to every customer with a contract expiring in 2027” — bulk apply
  • “What tags are most common on lost opportunities?” — segmentation analytics

​
comments

  • “Add a comment to the Globex opportunity: ‘They want the SOW by Friday’” — quick note logging
  • “Show me everything I’ve commented on this week” — personal activity recap

​
analytics

  • “What’s our average sales cycle by source?” — see the AI Analytics page for the full set
  • “Plot pipeline value by stage for the last 6 months” — time series with chart

​
pipelines

  • “What stages are in the Enterprise opportunity pipeline?” — pipeline + stage inspection
  • “Add a new stage ‘Security Review’ between Negotiation and Closed Won” — pipeline edit

​
communications (Twilio)

  • “Show me the call log with John Smith” — call history retrieval
  • “Send John an SMS: ‘Heads up, sending the proposal in 5 minutes’” — outbound SMS

​
web (search-the-internet agent)

  • “What’s Globex’s latest funding round?” — public web search for facts outside the CRM
  • “Find the LinkedIn profile of John Smith VP Sales at Stitch” — researcher-style lookup that doesn’t touch the CRM

​
The Automation Builder

A specific call-out because it’s a power user move: the automations agent uses the automation-builder prompt template to translate natural language into a saved automation rule. Example:
“When a contact’s source changes to anything containing ‘web’, round-robin-assign them to the sales team, tag them web-lead, and send them a welcome email after 1 hour.”
The agent:
  1. Parses the trigger event (contact_source_changed)
  2. Parses the trigger conditions (source contains 'web')
  3. Parses the sequence of actions (round_robin_assign, add_tag, send_email with a 1-hour delay)
  4. Drafts the rule and shows it to you for review
  5. Saves and activates on your confirmation
You can iterate from there — the agent can edit the draft if you say “actually delay the email by 2 hours instead of 1” or “add another step that creates a follow-up task after 3 days.”
The automation builder won’t activate a rule on the first try without your confirmation. You always get to eyeball the trigger + steps before they go live.