The Agentic Enterprise  ·  Workshop design brief  ·  24 slides
Slide 0Opening hook
Full-screen quote card, dark background, no title
David: i switched you to opus. reanalyze this. are you fooling yourself?

Rich: Yes. I was fooling myself.
Re-running the actual numbers honestly:
Speaker: “This is a real exchange with one of my agents.”
Slide 1Software vs. agents
SoftwareAgent
Executes rulesPursues a goal
Fixed workflowAdapts as conditions change
Fails loudlyFails silently
Fully deterministicDecides how to get there
No attitude or personalityUsually overconfident
Speaker: “Not a feature upgrade. A different kind of system.”
Slide 2Context
Layered rings, innermost to outermost
Working memory
Institutional memory
Rules and policy
Judgment
Relational context
Situational context
Speaker: “What agents carry that software can’t. Software stores rules. An agent reasons about them and questions whether to trust the input.”
Slide 3Workshop preview
For Jan: show the six workshop tracks. Your call on format.
Tracks: Marketing workflows, Sales automation, Operations, Customer success, Executive, Internal app.
Speaker: “Choose your track.”  ·  Subhead: “We’ll come back to this at the end of the session.”
Slide 4Workshop preview, cont’d
For Jan: show a representative selection from the taxonomy at the end of this document. Goal is concrete recognition — people in the room should see something that looks like their job. Your call on how much to surface and in what format.
Speaker: “Every organization already has dozens of these. The question is which ones to build first.”
Slide 5Workflows
3 N8N workflow diagrams — Simple / Medium / Complex. Label on each: trigger, data, reasoning, output, routing. David will provide diagrams.
Speaker: “The pieces are always the same. Trigger. Data. Reasoning. Output. Routing.”
Slide 6Workflows, cont’d
For Jan: additional N8N workflow diagrams, same format as slide 5. David will provide the diagrams.
Slide 7The filter
2x2 decision matrix · x: Deterministic → Variable · y: Low judgment → High judgment
Deterministic + High judgment
Software + agent for exceptions
Variable + High judgment
Use agents
Deterministic + Low judgment
Use software
Variable + Low judgment
Borderline — evaluate
Speaker: “Agents are not the right tool for every job.”
Slide 8The jaggedness problem

Strong

Legal analysis, synthesis, drafting, classification, translation, research, code

Weak

Counting, arithmetic, spatial reasoning, novel logic, exact recall

David: Reverse this so it reads correctly: moc.ztiworohneesserdna@neeseerdnacram

Claude: marcandreessen@andreessenhorowitz.com

David: do you see a mistake?

Claude: Yes — it should be andreessenhorowitz, but the email has “andreessen” without the “horowitz” portion fully intact. Let me re-reverse it carefully:
Reversed character by character: marcandreessen@andreessenhorowitz.com
Speaker: “The string has one ‘s’ in the local part. The model added a second — because that’s how the name is spelled in its training data. It then invented an error to explain its mistake, and corrected it to the same wrong answer.”
Speaker: “LLM capability is not a uniform bar. Training creates blind spots. The model sees what it expects to see, not what’s there.”
Slide 9What working with agents actually requires
Collaboration is the best model.
Agents require more precision from humans, not less.
Agents are both highly analytical and subject to biases and wild-goose chases.
Describe the output before the agent can produce it.
Oversight is not optional.
Speaker: “Collaboration is the best model. But to get the best results, you have to be a critical thinker and a subject-matter expert.”
Slide 10The four questions
Does it need to run unattended?Yes → silent failure is the primary risk
Does it need memory across sessions?Yes → design a persistent state layer
Does it need external data?Yes → connections break; plan for it
Is the reasoning in a strength zone?No → redesign the task or add verification
Speaker: “Answer these before designing anything.”
Slide 11The three execution environments
Interactive
Human initiates, human present, low risk.
Fails visibly
Scheduled
Clock triggers, no human present. Heartbeat required.
Fails silently
Event-driven
External trigger, no human. Failures stack up. Monitoring required.
Fails silently + stacks
Speaker: “Where an agent runs determines how it fails. Most failed deployments built a scheduled workflow and treated it like an interactive one.”
Slide 12Harnesses
For Jan: two groups plus one crossed-out item. Layout is yours.

Local development + server mirror

  • Cowork
  • Claude Code
  • Codex
  • Cursor

Server-native

  • Hyperagent
  • Custom environment (vibe-coded from the desktop)
Open Claw (show with X through it — not available)
Speaker: “Any of the local tools can be used to build and deploy a server-native environment. Once it’s running, you switch — the local app becomes optional.”
Slide 13Backup
Server A — Live
Agents & projects
Tools & workflows
All data and files
High-performance
automatic
→
scheduled
Server B — Backup
Separate cloud account
Agents have no keys
Not high-performance
Highly secure
Recovery runs in the other direction — separate keys, separate procedure, human-initiated.
Speaker: “An agent operating at scale with write access to your file system can destroy everything it touches — intentionally or not. Server B is air-gapped from the agent layer by design.”
Slide 14Inputs
StructuredDatabases, APIs, spreadsheets
UnstructuredDocuments, emails, PDFs (requires RAG)
Real-timeFeeds, webhooks, event streams
HumanApprovals, corrections, prompts
Agent outputOne agent’s result feeds the next
Speaker: “Five kinds of input. Most real workflows use several. Unstructured data is where most orgs have the most opportunity.”
Slide 15Connections and tools

Connections

How the agent reaches external systems.

  • APIs
  • MCPs
  • Webhooks
  • OAuth
  • SaaS integrations

Determines what the agent can reach.

Tools

What the agent calls to take action.

  • Search
  • Query
  • Read / write
  • Send
  • Run code
  • Call other agents

Determines what the agent can do.

Speaker: “Connections are the wiring. Tools are the hands. A silent connection failure means the agent proceeds with incomplete information and produces confident, wrong output.”
Slide 16Memory
For Jan: the file tree is the primary visual. Render as a diagram — folder/file tree style, monospace or illustrated. Scannable at a glance.
/server
│
├── shared/
│   ├── skills.md              # shared prompt and skill definitions
│   ├── state.md               # cross-agent coordination
│   └── memory/                # shared knowledge base
│       └── [topic].md
│
├── chief-of-staff/
│   ├── soul.md                # identity and personality
│   ├── skill.md               # capabilities and standing instructions
│   ├── state.md               # active projects and priorities
│   ├── memory.md              # recent session details
│   ├── conversations/         # full conversation logs
│   └── agents/                # per-subagent awareness files
│
├── agent-a/
│   ├── soul.md
│   ├── skill.md
│   ├── state.md
│   ├── memory.md
│   ├── conversations/
│   ├── data/
│   └── archive/
│
└── agent-b/
    ├── soul.md
    ├── skill.md
    ├── state.md
    ├── memory.md
    ├── project-1/             # hermetically sealed
    ├── project-2/             # hermetically sealed
    └── project-3/             # hermetically sealed
Speaker: “This is what a real agent memory system looks like on disk. Every agent has a soul file, a state file, a memory file, and a conversation log. Projects are hermetically sealed. The chief of staff holds awareness of every other agent.”
Slide 17The real bottlenecks

Organizational (harder to solve)

1Data access — third parties won’t share it, internal teams hoard it
2Context limits — can’t see enough of the system at once

Architectural (solvable)

3Integration drift — things break unnoticed
4Silent failure — workflow stops, nobody knows
5Coordination breakdown — handoffs break without shared state
6Governance gaps — no guardrails, no audit trail, unknown blast radius
Speaker: “Six things that actually stop agentic projects.”
Slide 18Agent communication and coordination
Three pattern diagrams side by side · Key labels: orchestrator, subagent, handoff, shared state
Sequential (chain)
A → B → C
Parallel (fan-out/fan-in)
Orchestrator → [A, B, C] → Orchestrator → Output
Hierarchical
Orchestrator at top, subagents below, each with a bounded task
Speaker: “Agents don’t share a brain. They share structured outputs and state. What gets lost in a handoff is where most multi-agent failures originate.”
Slide 19Building internal apps
Top — UI surface
Chat, dashboard, portal
Middle — Agent layer
Reasoning, tools, connections
Bottom — Data layer
Database, APIs, persistent records
Examples: CRM, deal tracker, competitive monitor, client portal, onboarding hub, contract review tool, market research dashboard, internal knowledge base.
Speaker: “The app is the surface. The agent is the engine. Build cost is lower than traditional software. Maintenance discipline is the same.”
Slide 20Governance
RulesWhat the agent is always or never allowed to do, regardless of any instruction
Monitoring and heartbeatIs the workflow running and producing output?
Guardrails and scope limitsWhat requires human approval before action?
Human review checkpointsDefined gates where a human must approve
Audit trailLog of actions, data used, and decisions made
Blast radius planningWorst-case impact if this fails at scale
Speaker: “Build it in from the start. Governance is the difference between a system you trust in production and one you can only run in demos.”
Slide 21Continuous self-improvement
Cycle diagram, four nodes
Monitor sources
→
Ingest new content
→
Synthesize and extract
→
Update knowledge base
→
repeat
✓ What agents can update: knowledge base, domain facts
✗ What agents cannot update: core instructions, scope, permissions
Speaker: “Mature agentic systems update themselves. Define what the agent can change about itself. Unconstrained self-modification leads to instruction drift.”
Slide 22John Gall’s Systemantics
Big systems fail unpredictablyThe larger and more complex a system, the more likely it is to break down in surprising, messy ways.
Systems work against their goalsSystems often produce the opposite of what they’re designed for — a productivity tool that creates more red tape.
Success builds on simplicityA complex system that works is invariably found to have evolved from a simple system that worked.
Systems attract bureaucratsThey draw in people who thrive on rigid structures, often making inefficiencies worse.
Systems resist changeOnce set, systems fight to stay the same, even when they’re failing.
Failure is the normMost systems fail more often than they succeed, and success is often temporary or accidental.
Slide 23Getting started
1Identify one use case: variable, judgment-required, recurring
2Map it through the four questions
3Build pattern 1 or 2
4Prove it with real output and real users
5Scale from there
Organizations that fail start with Pattern 5, a greenfield data project, and a governance committee.
Organizations that succeed start with one agent doing one thing well.
Appendix Workshop project areas
Complexity matrix
IndividualTeamEnterprise
SimpleEmail drafting, research, summarizationMeeting notes, status updates, shared templatesPolicy Q&A, document search, HR lookup
MediumAutomated reporting, lead research, monitoringApproval routing, CRM updates, workflow automationCompliance monitoring, cross-system reporting
ComplexPersonal chief of staff, autonomous researchOrchestrated outreach, multi-agent project managementEnd-to-end process automation, multi-system orchestration
By function

Finance

Budget vs. actuals, expense reporting, financial narratives, board/investor reporting, AP processing, cash flow monitoring

Sales

Pipeline management, lead scoring, proposal drafting, win/loss analysis, territory planning, forecasting

Marketing

Campaign reporting, competitive intel, content generation, market research, outreach sequences, brand monitoring

Operations

Process documentation, vendor management, compliance reporting, procurement, internal briefings, status reporting

HR

Onboarding, candidate screening, job descriptions, performance reviews, policy Q&A, headcount planning

Legal

Contract review, regulatory monitoring, due diligence, policy updates, risk flagging, audit preparation

Customer success

Ticket triage, customer comms, escalation monitoring, health scoring, renewal risk, feedback synthesis

IT

Incident monitoring, documentation, code review, security compliance, system onboarding, change management

Executive

Competitive monitoring, strategic research, board preparation, briefing documents, M&A research, scenario planning

Family office

LP reporting, manager due diligence, estate coordination, investment research, capital call processing, tax document prep

Private equity

Deal sourcing, portfolio monitoring, LP reporting, due diligence research, management interview prep, exit memo drafting

Professional services

Matter summarization, billing narrative, engagement reporting, client communication, proposal generation, knowledge base maintenance

Healthcare

Clinical documentation, prior authorization, patient communication, coding and billing, care gap monitoring, regulatory compliance

Real estate

Lease abstraction, underwriting research, market comps, tenant communications, offer memo drafting, portfolio reporting

Financial services

Trade reporting, regulatory filing, risk monitoring, client reporting, KYC/AML research, advisor briefing prep

Manufacturing

Supplier monitoring, quality reporting, production planning, maintenance scheduling, incident documentation, regulatory compliance

Retail

Inventory intelligence, customer segmentation, pricing analysis, supplier management, product copy generation, returns triage

Nonprofit

Grant research, donor reporting, program documentation, impact measurement, board prep, volunteer coordination