What stops the pack from giving wrong answers?

Every agent loads KB grounding before responding. KB Answerer cites doc URLs per claim. If a question isn't in the KB, the agent escalates instead of guessing. The whole architecture is doc-grounded — no 'best guess' generation. Customers either get a sourced answer or get a human.

Can it handle multi-language support?

Yes. Each customer's preferred language goes in their account-context memory. KB Answerer can use language-specific KBs or translate from one canonical KB. Escalation Router routes to language-matched humans where available.

How does this work with our existing helpdesk?

MCP connectors for major helpdesks — Zendesk, Intercom, Front, HubSpot Service. Tickets flow in and out, status syncs, agent comments flow into customer memory. The pack doesn't replace your helpdesk; it adds an agent layer that operates inside it.

What about sentiment and tone matching?

Account Context tracks customer sentiment over time. Responses get written in a tone configured per customer or per situation. A frustrated long-time customer gets different language than a first-time user with a quick how-to question.

Can it actually defuse angry customers?

Better question is: should it. Most teams configure 'urgent' or 'angry' triage to escalate immediately to humans, not auto-respond. The pack handles volume tickets where the answer is in the KB; humans handle the moments that need judgment. Both work in the same system.

PACKWOLF is in closed beta with small cohorts; pricing is set per cohort while we tune the pack to your channels and KB. General-availability pricing lands when the beta closes.

Use case · Support

Compose a specialist agent for every layer of support.

One agent triages inbound tickets, another answers from your KB, another tracks per-customer history, another routes escalations, another surfaces product patterns — every agent KB-grounded, every escalation traced, your humans handle the moments that need judgment.

Best fit for · Support teams running tier-1 deflection across multiple channels with a real knowledge base and escalation rules

Request beta access See an example pack

vs Lindy →

packwolf.app · Support inbox

Live

An example support pack working a ticket. Yours can look different — you define which agents own triage, KB answers, escalation, and pattern reporting.

What the pack does

Five concrete tasks.

01
Specialist agents for every layer of support.
Compose the pack you need: a Triage agent for inbound classification, a KB Answerer for doc-grounded responses, an Account-context agent for per-customer history, an Escalation Router, a Voice-of-Customer agent for pattern detection. Add or drop specialists as your channel mix evolves.
02
KB grounding loads into every agent.
A KB Custodian (or your KB memory) holds the doc index, citation rules, and the do-not-fabricate guardrail. Every other agent — Triage, KB Answerer, Escalation Router, Voice of Customer — loads it before responding. If a question isn't in the KB, the agent escalates instead of guessing.
03
Triage every ticket in seconds.
Triage agent classifies (billing / technical / account / urgent), pulls account context, and decides: KB-answerable, needs-human, or hybrid (KB answer drafted for human review). Tickets stop sitting in queues for hours waiting for first-touch.
04
Remember the customer.
Account-context agent loads everything: prior tickets, plan tier, last conversation tone, sentiment trend. Customers don't repeat themselves; the pack doesn't repeat the customer. Long-time frustrated users get different language than first-time how-to questions.
05
Escalate cleanly, surface patterns.
Escalation Router applies your hand-off rules — the right human gets the ticket with full context. Voice of Customer aggregates patterns weekly: which features generate tickets, which docs need updates, which cohorts churn. The pack feeds product, not just support.

How a team uses this

Setup once, then watch it run.

Concrete operator setup, the phases the pack moves through, and where you stay in the loop.

Connect your support inbox (Gmail or shared mailbox) as an integration with read + send scopes.

Connect Telegram so the pack can ping you when a reply needs a human call.

Seed the playbook by pasting your three best past replies, the pack treats them as procedural memory.

Set tier policy: auto-send for tier-1, draft-only for tier-2, escalate-on-uncertainty.

Phase 1

Watch

A scheduled heartbeat polls the inbox. New messages stream in, the Triage agent reads tone, urgency, and topic against the playbook.

Phase 2

Draft

The Drafter pulls the matching playbook entry from durable memory and writes a reply. Voice agent checks tone, fixes drift before anything leaves.

Phase 3

Escalate?

If the question needs facts only the human knows, the pack pings you on Telegram with the original message and a proposed reply. You answer in one tap.

Phase 4

Send + learn

Sender ships the final reply through the same integration. Logger updates the playbook so the next identical question is one less escalation.

An example pack + workflow

Your pack, your workflow.

Workflows are markdown that reference the roles in your pack. Below is one example shape - yours can have different agents, different steps, different cadence.

Step 01Seconds
Inbound triage
ReceivesCustomer's question + channel
TRTriageACContext
Triage classifies (billing / technical / account / urgent). Account Context loads per-customer history. Triage decides routing: KB / human / hybrid.
Hands offClassified ticket + routing decision
Step 02Seconds
KB answer (or hybrid)
ReceivesClassified ticket
ANAnswererKBKB
KB Answerer responds from your docs only, citing URLs per claim. Anything outside the KB skips to escalation; hybrid drafts wait for human review.
Hands offDrafted answer with citations OR escalation flag
Step 03Within SLA
Escalation routing
ReceivesEscalation flag
ERRouterEPPolicy
Escalation Router applies your hand-off rules. Hand-off package: customer history, classified intent, what the pack tried, KB references checked, SLA clock state.
Hands offTicket package to right human
Step 04On close
Resolution + memory write
ReceivesClosed ticket
ACContext
Key facts write to per-customer memory: resolved issue, preference expressed, sentiment, escalation cost. Customer profile stays current.
Hands offUpdated customer profile + closed ticket
Step 05Weekly
Pattern detection
ReceivesWeek's resolutions
VCVoCQAQA
Voice of Customer aggregates patterns. QA Analyst pulls resolution quality, time-to-resolve. Reports go to product + docs teams.
Hands offProduct feedback report → KB updates next cycle

Example workflow · 5 steps · per-ticket handoff chain

Workflow · markdownExample inbound ticket flow

# Example inbound ticket flow
# Workflows are markdown — yours can swap agents,
# add channels, tighten SLAs, or define a different
# escalation matrix.

Match: tasks labelled "support-ticket"
Required tools: http_request (KB + ticketing via MCP), email_send
Required skills: customer-empathy, escalation-policy

## KB grounding — always-on
Every agent loads from your KB memory before responding: doc
index, citation rules, do-not-fabricate guardrail. No agent
answers without grounding.

## Step 1 — Triage (seconds)
Triage classifies: billing / technical / account / urgent.
Account-context agent loads per-customer history. Triage
decides routing: KB / human / hybrid.

## Step 2 — KB answer (when applicable)
KB Answerer responds from your docs only, cites doc URLs per
claim. Anything outside the KB skips to escalation.

## Step 3 — Hybrid (when applicable)
KB Answerer drafts an answer for human review. Customer
doesn't see the draft until a support agent approves or edits.

## Step 4 — Escalate (when needed)
Escalation Router applies your routing rules. Hand-off package:
customer history, classified intent, what the pack tried, KB
references checked, SLA clock state.

## Step 5 — Resolution + memory write
On close, key facts write to per-customer memory: resolved
issue, preference expressed, sentiment, escalation cost.

## Step 6 — Pattern detection (weekly, background)
Voice of Customer aggregates patterns. QA Analyst pulls
resolution quality, time-to-resolve, and feeds back to
product + docs teams.

Approvals: Support Lead signs off on escalation policy and
auto-respond rules monthly.

Built on

CapabilityMemory CapabilityWorkflows CapabilityTools CapabilityApprovals

The ROI

Tier-1 deflection on volume tickets — composable, KB-grounded, with a product feedback loop that didn't exist before.

Common questions

Things teams actually ask.

No. Support teams shape differently — some need a dedicated billing agent, some a separate Spanish-language pack, some want a churn-prevention agent that only watches at-risk accounts. PACKWOLF lets you compose the specialists you actually need. The example here is one shape, not the shape.

Related use cases

Use caseMarketing automationWhere the KB content originally gets written Use caseSales prospectingSame triage logic, sales side Use caseCompliance monitoringAudit trail per resolution

Run this pack on your team's work.

Closed-beta cohorts are small. Tell us about your work and we'll configure the pack for what you actually do.

Request beta access

Compose a specialist agent for every layer of support.

Specialist agents for every layer of support.

KB grounding loads into every agent.

Triage every ticket in seconds.

Remember the customer.

Escalate cleanly, surface patterns.

Watch

Draft

Escalate?

Send + learn

Inbound triage

KB answer (or hybrid)

Escalation routing

Resolution + memory write

Pattern detection

Run this pack on your team's work.