Skip to content
Skip to main content
PLATFORM · LAYER 04 · THE AGENTS

Every exec asks for
an AI agent.
Most are three layers below the maturity
needed to support one.

We don't say that to be condescending. We say it because we've seen what happens when companies skip ahead. The Agents layer is the autonomous workforce that runs on top of your kernel — so the work your team doesn't want to do stops requiring your team to do it.

Book the diagnostic →See Layer 03 · Kernel
90 minutes. No slides. Just the map.

Eight steps to an incident.

We've fixed several of these. We'd rather build the Kernel first and ship agents that survive in production.

  1. 01
    CEO reads about AI
    Asks marketing for "an AI agent for customers."
  2. 02
    Vendor builds the agent
    Demo looks great in a sandbox.
  3. 03
    Agent ships to production
    On top of an L2 kernel — the substrate isn't ready.
  4. 04
    Works for about a week
    Handles shallow queries. Confidence builds.
  5. 05
    A customer asks a real question
    One the kernel has no clean answer for.
  6. 06
    The agent hallucinates
    Confidently. With brand voice intact.
  7. 07
    Customer quotes the answer back
    In a meeting. In a support ticket. On LinkedIn.
  8. 08
    Incident
    Trust collapses. The next AI initiative starts with a credibility deficit.

Three flavors of agent. All real. All useful.

Most companies start with co-pilots, get value, and stop. We push into automations and autonomous teammates because a properly wired kernel can support it.

FLAVOR 01
WHERE MOST COMPANIES STOP

Co-pilots

Human in the loop on every output.
Agents that sit next to a human and accelerate them. Sales reps writing follow-ups. CS reps responding to tickets. Marketers drafting content briefs.
  • Sales follow-up drafts
  • Ticket triage + suggested replies
  • Content brief assembly
FLAVOR 02
WHERE A CLEAN KERNEL UNLOCKS IT

Automations

End-to-end. No human required.
Agents that handle defined work without a human in the loop. Lead routing. Data hygiene. Multi-step workflows that used to require a Zapier graveyard.
  • Lead routing
  • Data hygiene jobs
  • Multi-step workflows
FLAVOR 03
WHERE THE COMPOUNDING STARTS

Autonomous teammates

Owns a category. Reports back.
Agents that own a category of work, make decisions inside guardrails, and report back. The hardest to build well. The most valuable when they work.
  • Category ownership
  • Decisions inside guardrails
  • Reports back to the team

Four deliverables. One stack.

Most engagements include all four; the AI Maturity Program is the multi-quarter version.

DELIVERABLE · 01

Breeze Agent Library

Productized Breeze agents tuned to your specific operational patterns.

  • Sales co-pilot agentdrafts follow-ups, summarizes call notes, surfaces next-best-action
  • CS co-pilot agentresponds to tickets, summarizes customer history, suggests solutions
  • Marketing co-pilot agentdrafts briefs, repurposes content, applies CLEAR scores
  • Prospecting Agent configtop-of-funnel buyer-facing work
  • Customer Agent configwith the governance posture to put it in production
  • Custom Breeze agentsfor operational patterns specific to your business
Each agent ships with documentation, training, audit logging, and a defined success metric.
Breeze Agent Library →
DELIVERABLE · 02

Custom MCP Servers

For exposing your kernel cleanly to agents inside or outside HubSpot.

  • Protocol surfaceread, write, search, resource, and tool endpoints — specific to your data model
  • Auth + scope governanceOAuth or signed tokens. Per-agent, per-operation scoping.
  • Audit log infrastructureevery call, every response, every actor, every timestamp. Queryable.
  • Rate limits + output controlsand rollback for write operations.
  • MCP integrationClaude, ChatGPT, custom agents, and partner systems — as your governance permits.
Your kernel becomes legible to any agent you authorize — no rebuilds.
Custom MCP Servers →
DELIVERABLE · 03
FLAGSHIP

AI Maturity Program

The flagship engagement. Multi-quarter. Moves your company from L2 toward L4 deliberately.

  • Q1 · Kernel workL2→L3 transition. Data model, workflows, reporting, Breeze, brand-as-prompt.
  • Q2 · Protocol installCompany MCP surface stood up. Auth + governance wired. First wave of internal agents (3–5 co-pilots).
  • Q3 · Agent portfolioCustomer or partner-facing agent. Governance matured. Output controls hardened.
  • Q4+ · Operating partnershipWe tune, expand the portfolio, and operate with you as you grow capability internally.
Most clients ship 3–5 agents in the first 90 days, plus the substrate to ship more without us.
AI Maturity Program →
DELIVERABLE · 04

Agent Governance & Audit

The discipline that makes agents safe to scale.

  • Governance policy docsspecific to your agents: what each can do, what it can't, where humans stay in the loop.
  • Audit log infrastructureevery action logged, searchable, retained per policy.
  • Output control frameworkschema enforcement on writes. PII redaction. Brand-voice consistency checks.
  • Incident responsewhen an agent goes off-script (it happens), the playbook to investigate, contain, roll back.
  • Versioning + rollbackfor prompt and tool changes.
  • Cost monitoring + rate limitingto prevent runaway spend.
  • Team trainingon how to operate, monitor, and improve the agent portfolio.
Wired before launch, not after the first incident.
Agent Governance & Audit →

Why agents fail in production. And how we prevent it.

We address each of these explicitly in every Agents engagement. Not as a checklist; as the design.

Failure mode
What goes wrong
How we prevent it
Vague scope
"Help customers." Defines nothing. Agent fails at everything.
Scope each agent to a specific category of question or task. "Answer billing-status questions." "Generate change-of-plan quotes."
No real substrate
Agent has nothing accurate to read from. Hallucinates.
Build the Kernel first. Audit substrate readiness before deploying any agent.
Default voice
Agent sounds like every other company's agent. Customers feel like the company outsourced.
Voice layer feeds the agent's prompt. Brand-as-prompt rules loaded as system context.
No escalation path
Customers can't reach a human. Agent becomes the wall instead of the front door.
Clean escalation as a first-class flow. One click. Full context preserved for the human picking up.
No governance
No audit log, no output controls, no rate limits. Something goes wrong. there's no way to investigate or roll back.
Governance layer wired before launch, not after the first incident.

Wherever the work happens.

A sales co-pilot belongs in the rep's inbox. A data-hygiene automation belongs in the workflow engine. A customer-facing agent belongs at the Surface, backed by the Kernel. We don't pick one venue. The protocol is what makes deploying across venues safe and scalable.

INSIDE THE KERNEL
HubSpot · Breeze
Native HubSpot agents. First-class kernel access.
PROTOCOL LAYER
Custom MCP server
Your own protocol surface. Per-agent, per-operation scopes.
EXTERNAL · GOVERNED
Claude · ChatGPT
External LLMs connected to your kernel under governance.
SURFACE LAYER
Embedded customer UI
Agent at the surface, backed by the kernel.
TEAM SURFACE
Slack bot
Internal team interface. Co-pilot in the channel.
AUTONOMOUS
Cron · scheduled
Nightly scans. Quarterly clean-ups. Always-on jobs.
The protocol →
6mo+
AGENT UPTIME · IN PRODUCTION, NOT QUIETLY DISAPPEARING
L2→L3
IN 90 DAYS · MATURITY TRANSITION
3–5
AGENTS SHIPPED · FIRST 90 DAYS
↓ cost
PER RESOLUTION · INCL. HUMAN ESCALATIONS

Numbers your CFO will recognize.

Deflection rate
Work the agents handle that used to require a human.
Trust meter
Repeat-use rate. Customers and team members coming back to the agent.
Escalation quality
When the agent escalates, is the human's job easier because of the context it passed?
Time-to-answer
End-to-end, including escalations.

Read from. Render in. Sound like.

LAYER · 01
The Surface
The Agents render here — so customers, partners, and your team can talk to them.
→
LAYER · 02
The Voice
The Agents sound like this — recognizably your company, not a generic assistant.
→
LAYER · 03
The Kernel
The Agents read from here — so they have something real to call.
→
LAYER · .
The Protocol
The seam that lets it all work. Company MCP is what makes deploying safely possible.
→

The diagnostic is pre-loaded.

You walk away with a ranked list of agent opportunities, a written map of what your kernel needs before each one can run reliably, and an honest read on whether agents are the right next move — or whether kernel work needs to come first.

  • Audited your existing automations and AI experiments
  • Read your AI strategy doc if you have one
  • Identified the 3 highest-value agent candidates in your kernel
  • Built a maturity-level read with specific evidence
  • Assessed your kernel's readiness to support each candidate

Things people ask before signing.

Do we need a clean kernel before we can ship a single agent?

Not always. We can ship a co-pilot agent against a messy L2 kernel. they're forgiving because a human is in the loop. Customer-facing agents and autonomous teammates need at least L3.

What's the relationship between Breeze and a custom MCP server?

Breeze is HubSpot's native agent layer, sitting inside the kernel. A custom MCP server is your protocol surface for external agents. Most clients have both.

Who owns the agents and the audit log after handoff?

You do. Fully. The Breeze configs, the MCP server, the prompt library, the governance docs, the audit infrastructure — all on your HubSpot, your data, your governance.

How long until the first agent is in production?

If kernel readiness checks out, the first co-pilot ships in 4–6 weeks. Customer-facing agents take longer — usually 8–12 weeks.

What if an agent goes off-script in production?

It happens. The Governance deliverable includes the incident-response playbook: pause the agent, query the audit log, reproduce the failure, roll back.

Pricing?

Three lines on the engagement: implementation (one-time, scoped per deliverable), Coworker platform access (recurring), ongoing services (recurring or project).

DIAGNOSTIC · 90 MINUTES · NO SLIDES

Map your company's intelligence in 90 minutes.

No slides. Just the map. We'll whiteboard what your company already knows, what it's forgetting, and which layer to build first.

Book the diagnostic

Yours to keep — signed or not.