🎮 The Next Input — Issue #120

In partnership with

⚡ The Briefing — 60 sec

• Anthropic launches interactive Claude apps for Slack and workplace tools
  At this point it’s less about what these tools can do—and more about what they won’t be able to do in the future.

• Davos debates AI, energy, geopolitics—and somehow Greenland
  Pardon me sir, you left your Greenland in my OpenAI-branded Louis Vuitton swag bag.

• Australian jobs most at risk from AI as CEOs weigh in on the future of work
  Funny how none of those CEOs see their jobs in jeopardy. Whole different conversation.

🛠️ The Playbook — The Workplace Agent Mesh

Mission Turn Slack into a controlled agent layer that triages, drafts, and executes work across your tools—without “bot sprawl.” Difficulty Advanced Build time 3 hours ROI 6 hrs/week saved → 2 hrs/week (measured via Slack + ticket timestamps) with fewer dropped handoffs.

0) Why This Matters

Agents inside workplace tools are inevitable—Slack is becoming the control plane. The risk is letting 15 “helpful bots” quietly turn into 15 unowned systems making untracked changes. This mesh gives you speed and accountability: who did what, where, and under which rules.

1) Architecture

2) Workflow

1. Request intake (Slack): User posts in #ops-requests or runs /agent help with the request.

2. Scope check: Router labels the request type: research, draft, ops-change, finance, people, unknown.

3. Decision gate:

   • If unknown → ask one clarification question and stop.

   • If finance or people → default to draft-only + require approval.

4. Routing: GPT-5-mini selects the correct agent (single agent only).

5. Execution: Claude 4.5 Sonnet agent completes one task: draft, summarise, extract, or propose steps.

6. Write gate (critical): If the task touches a system change (ticket updates, record edits, deployments):

   • If confidence < 0.8 → do not write; return a proposed patch + ask for approval.

   • If confidence ≥ 0.8 → still require explicit /approve for writes.

7. Action + log: Once approved, the bridge executes and logs: input, output, diff, actor, timestamp.

8. Close loop: Agent posts a final Slack summary with: what changed, where, and rollback path.

3) Example Prompts

Router (GPT-5-mini)

Classify and route this Slack request.

Return JSON:
{
  "category": "research|draft|ops-change|finance|people|unknown",
  "best_agent": "name",
  "needs_clarification": true/false,
  "one_question_if_needed": "",
  "requires_approval": true/false,
  "risk_level": "low|medium|high"
}

Rules:
- If category is finance or people, requires_approval = true.
- If uncertain, category = unknown and ask one question.

Executor (Claude 4.5 Sonnet)

You are a scoped workplace agent. Complete ONE task only.

Task: {task}
Constraints:
- Do not write to external systems.
- If information is missing, ask one question and stop.
Output:
1) Result
2) Assumptions (if any)
3) Next action + whether approval is required
Tone: concise, operator-friendly.

Eval (Claude 4.5 Haiku)

Evaluate the agent output for:
- scope compliance
- missing steps
- risky assumptions
Return: PASS / FLAG / FAIL
If FLAG/FAIL, provide the minimal fix.

4) Guardrails

• Draft-only by default. Writes require explicit approval.

• One agent, one job. No “general assistant” agents.

• No silent permissions. Any new permission requires a human owner sign-off.

• Immutable audit trail for every action and diff.

• Hard block on sensitive domains (HR, health, legal) unless configured + approved.

5) Pilot Rollout — 3 hours

1. Create Slack channels: #agent-requests, #agent-audit, #agent-incidents.

2. Stand up 3 agents only: Research, Comms Drafting, Ops Triage.

3. Implement routing + eval prompts and force draft-only.

4. Run 20 real requests through the system.

5. Success criteria: ≥80% correct routing, ≤2 minutes to first draft, 0 unapproved writes, rollback tested once.

6. Expand to one write-capable workflow with /approve (e.g., Jira ticket updates).

6) Metrics

• Time to first useful output: 10 min → 2 min (Slack timestamps)

• Correct routing rate: 60% → 85% (manual sample of 20/week)

• Unapproved write attempts: baseline unknown → 0 (OPA policy logs)

• Task completion time (ops triage): 45 min → 20 min (ticket lifecycle)

• Repeat failure rate: 5/week → 1/week (Langfuse error tags)

Pro Tip: If an agent asks for “just one more permission,” that’s not progress—that’s scope creep.

🎯 The Arsenal — Tools & Prompts

• Claude Apps for Slack · Build native interactive agents inside Slack · https://www.anthropic.com

• Open Policy Agent · Enforce approvals, scopes, and permissions as code · https://www.openpolicyagent.org

• Langfuse · Trace every run, catch drift, and debug failures fast · https://langfuse.com

• Temporal · Run durable agent workflows that don’t break mid-task · https://temporal.io

Copy-paste prompt block:

You are a workplace agent operating under strict controls.

Rules:
- Draft-only unless explicitly approved.
- One task at a time.
- If unclear, ask ONE question and stop.
- Never assume permissions.
- Always return: result, assumptions, next action, rollback.

If the task touches money, people, or policy: require approval.

💡 Free Office Hours

Want help implementing anything? Book a free 15-minute Office Hours slot—no sales pitch, just workflows solved.

→ https://calendly.com/aaron-cylentis/the-next-input-office-hours

Free, private email that puts your privacy first

A private inbox doesn’t have to come with a price tag—or a catch. Proton Mail’s free plan gives you the privacy and security you expect, without selling your data or showing you ads.

Built by scientists and privacy advocates, Proton Mail uses end-to-end encryption to keep your conversations secure. No scanning. No targeting. No creepy promotions.

With Proton, you’re not the product — you’re in control.

Start for free. Upgrade anytime. Stay private always.

Get free private email

🕹️ Game Over

More agents doesn’t mean more chaos—unless you let it.

— Aaron Automating the boring. Amplifying the brilliant.

Subscribe: https://cylentisai.beehiiv.com/subscribe