Where agents win for cloud cost

WHERE AGENTS WIN:
  Multi-step research questions across data sources
  Cross-referencing audit + cost + recommendation data
  Drafting written artifacts (memos, tickets, postmortems)
  Natural-language exploration of unfamiliar data
  Repeated tasks where prompts can be saved and reused

WHERE AGENTS DO NOT WIN:
  Already-automated workflows (schedules, anomaly detection)
  Dashboard browsing (UI is visually faster)
  Hard math / forecasting (use platform models)
  High-stakes decisions (human judgment required)
  Bulk operations (UI bulk actions are faster)

QUESTION: "Why is team-platform spend up 30% this month?"

WITHOUT AGENT (manual):
  Open ZopNight cost dashboard
  Filter to team-platform
  Toggle current vs last month
  Drill into top growing resources
  Cross-reference with recommendations (anything not applied?)
  Cross-reference with audit log (any schedule changes?)
  15 minutes of clicking through different surfaces

WITH AGENT (MCP-connected):
  Engineer: "Why is team-platform up 30% this month?"
  Agent: chains four MCP calls:
    1. get_costs(team=platform, period=current_month, last_month)
    2. list_resources(team=platform, sort=growth_rate, top=5)
    3. list_recommendations(team=platform, status=open)
    4. list_audit_logs(team=platform, filter=schedule_changes, 7d)
  Synthesizes: "EC2 grew $4,200 driven by 3 new prod-payments
                instances launched 2026-05-03 by jane@platform.
                No schedule changes. No open recs would have
                prevented it. Investigating with @jane."

  90 seconds end-to-end.

INCIDENT: cost anomaly fired at 14:00 today
ENGINEER needs to write comms for 3 audiences (internal eng,
finance, leadership)

WITHOUT AGENT:
  Engineer writes 3 messages from scratch
  Each requires switching context + tone
  30-45 minutes for the three

WITH AGENT:
  Engineer: "Draft incident comms for the cost spike at 14:00.
            Three audiences: internal eng (technical), finance
            (impact-focused), leadership (high-level summary)."
  Agent: pulls anomaly details from MCP, drafts 3 messages with
         appropriate tone and detail level for each audience
  Engineer reviews, edits, sends
  10 minutes total

AFTER a cost incident:
  Agent generates the postmortem skeleton:
    Timeline from audit logs
    Resources affected from resource summary
    Cost impact from anomaly delta
    Suggested action items based on root cause pattern
  Engineer fills in the analysis and validates

  TIME SAVED: 1 hour of formatting + assembly
  Engineer focuses on judgment (lessons, action items)
  not data collection

NEW engineer joins the team. Doesn't know:
  Which clusters are owned by which team
  What the normal cost baseline is
  What recent incidents looked like
  Where the schedules live and what they do
  Who to ask about specific resources

Without agent:
  Senior engineers field "where do I find X?" questions for weeks

With agent:
  New engineer asks agent any of those questions
  Agent answers from MCP data
  Cuts senior-engineer interrupt time by 80%+
  New engineer ramps faster

DASHBOARD BROWSING:
  A human flicking through a chart is faster than asking the agent.
  Visual pattern recognition beats text description of trends.
  When the data fits on one chart, look at the chart.

REPEATED ATOMIC ACTIONS:
  Killing 100 idle resources is a bulk action in ZopNight UI.
  Agent + human-in-the-loop = slower than the bulk feature.

HARD MATH:
  Forecasting, anomaly statistics, budget projection: use the
  platform's models. Don't ask the LLM to do arithmetic; it will
  approximate and sometimes be wrong.

HIGH-STAKES DECISIONS:
  Should we sign a $1M reservation? Agent can summarize the data.
  Decision is human.

REAL-TIME OPS:
  During an active incident, the dashboard + on-call runbook are
  faster than chatting with an agent.

ANYTHING THE UI DOES IN ONE CLICK:
  If the dashboard has the answer in one click, click the dashboard.
  Don't make the agent earn its keep on trivial work.

AGENT-FRIENDLY PROFILE:
  Multi-step
  Multi-source
  Written output
  Non-trivial synthesis
  Repeated (worth saving as a prompt)

EXAMPLES of work matching this profile:
  Weekly cost review prep
  Incident postmortem
  Quarterly trend summary for leadership
  New-engineer onboarding answers
  "How does team A spend differ from team B?"
  Drafting a chargeback explanation
  Pre-meeting briefing memo

ANTI-PATTERN                              WHY IT FAILS
──────────────────────────────────────────────────────────────────
Asking the agent for a single number       Open the dashboard;
                                          it's one click

Asking the agent to do arithmetic           LLMs approximate; use
                                          platform forecasting

Asking the agent before reading the         Often the answer is in
incident runbook                            the runbook; faster to
                                          read it

Treating the agent as the primary           UI is faster for known
interface (over the UI)                     workflows; agent for
                                          novel queries

"Let me ask the agent" as default           Often "let me look at
                                          the dashboard" is faster

Forwarding agent output without              Verify; agents can
verification                                hallucinate (especially
                                          numbers)

WORKFLOWS THAT TEND TO RECUR:
  Weekly cost summary by team
  Anomaly drill-down
  Postmortem skeleton
  Onboarding queries
  Recommendation review

ENGINEER OPENS Cursor with MCP connected:

Engineer: /weekly_cost_summary "team-platform"

Agent (chains five MCP calls):
  1. get_costs(team=platform, last_7_days)
  2. get_cost_trends(team=platform, compare=prior_7_days)
  3. get_recommendation_summary(team=platform)
  4. list_audit_logs(team=platform, filter=material_changes)
  5. list_resources(team=platform, sort=growth_rate, top=5)

Agent synthesizes (output to clipboard):

  "Weekly cost summary: team-platform (week of 2026-05-15):

  TOTAL: $12,400 this week, -3% WoW

  Top changes:
    EC2: +$200 (new instances for feature launch)
    EBS: -$340 (cleanup from last Friday)
    RDS: stable

  Recommendations:
    2 open totaling $480/mo unrealized savings
    Most impactful: right-size i-0abc → $300/mo

  Material changes (audit):
    1 schedule disabled 2 days ago by @bob (review)

  Top 5 growing resources:
    [list with dollar deltas]

  Action items for the week:
    Review @bob's schedule change
    Apply right-sizing on i-0abc"

Engineer pastes into team's weekly thread on Slack.

ELAPSED: 30 seconds of engineer time

TASK 1: __________
  Multi-source? __________
  Written output? __________
  Repeated? __________
  AGENT FIT: yes / no
  WHY: __________

TASK 2: __________
  Multi-source? __________
  Written output? __________
  Repeated? __________
  AGENT FIT: yes / no
  WHY: __________

TASK 3: __________
  Multi-source? __________
  Written output? __________
  Repeated? __________
  AGENT FIT: yes / no
  WHY: __________

For agent-fit tasks: save the prompt as a team skill (M6.4).
For UI-fit tasks: bookmark the dashboard view.