Use Cases + Implementation

How do you reduce the cost of support agents with model routing?

Support agent cost is loop cost: every step resends context, every tool call returns data that joins the prompt, and every retry repeats both. Route routine turns and tool follow-ups to cheap tiers, reserve strong models for escalation decisions and hard cases, cap steps and tool calls per conversation, and escalate instead of retrying when a step fails twice.

Key facts

Router modelplugsky-fusion escalates per step across tiers (live)
AgentsMemory, tools and orchestration are live
Function callingLive for lookups, actions and escalation triggers
Loop controlsStep, tool-call and retry caps in your orchestrator
PricingFlat monthly self-serve plans with no per-token charges on self-serve
Free tierplugsky-micro and plugsky-lite on the free plan, no card required
GovernanceScoped keys, RBAC and audit logs per agent action
RoadmapThe moderation endpoint is coming soon; apply your own filters today

TL;DR

  • Agent cost is loop cost: steps, tool results and retries compound.
  • Run routine turns and tool follow-ups on cheap tiers.
  • Escalate decisions and hard cases, not every step.
  • Cap steps, tool calls and retries per conversation.
  • Escalating to a human is often cheaper than another retry.

How it works, step by step

  1. Instrument steps per conversation, tool calls per step and retries per tool.
  2. Route triage, routine replies and tool follow-ups to cheap tiers.
  3. Escalate reasoning-heavy or sensitive steps to strong models.
  4. Return compact tool results and strip fields the agent does not need.
  5. Cap steps and tool calls per conversation; escalate on cap or double failure.
  6. Cache conversation summaries instead of replaying full history each step.
  7. Track cost per resolved conversation alongside resolution and escalation rates.
1Instrument stepsper conversation,tool calls per step2Route triage,routine replies andtool follow-ups to3Escalatereasoning-heavy orsensitive steps to4Return compact toolresults and stripfields the agent5Cap steps and toolcalls perconversation;6Cache conversationsummaries insteadof replaying full

Try it yourself

Open the AI agent cost calculator →

Where support-agent spend goes

An agent resolves a ticket through a loop: read the case, choose a tool, read the result, decide again. Steps multiply context, tool results add tokens, and failures add iterations. The dominant costs are usually replayed history, strong models used for mechanical steps, and retry loops that repeat the same failing call.

Each has a direct fix: keep a rolling summary instead of full history, match the tier to the step's difficulty, and convert repeated failures into escalation rather than another attempt.

Tool calls and context budgets

Tool output is where context grows fastest. An order lookup that returns every field bloats the prompt for every subsequent step, so return compact, purpose-built payloads:

  • Return only the fields the decision needs.
  • Summarise lists instead of pasting them.
  • Cache stable customer data for the conversation rather than refetching it.
  • Cap tool calls per step and per conversation, and log which tools consume the most turns.

Then route. Lookups, status checks and response drafting run cheaply; escalation decisions, policy interpretation and conflict resolution deserve stronger models.

Escalation as a cost control

Support agents fail expensively when they keep trying. A second failure on the same tool or the same intent should trigger escalation, not a third attempt with a reworded prompt. A human resolving the ticket in two minutes is cheaper than ten automated attempts and a frustrated customer.

Measure cost per resolved conversation, steps per resolution, tool failure rate and escalation rate. Rising steps with flat resolution is the signature of a routing or prompt problem. Start on the free plan with plugsky-micro and plugsky-lite for triage and drafts, then use the 14-day full-access trial for strong models on escalated cases; the live pricing page lists plans.

Honest comparison

Support agent choiceRouted agentStrong model every stepCheap model every step
Triage and repliesCheap tierFrontier price per stepAdequate
Tool follow-upsCheap tier with compact resultsExpensive context growthAdequate
Escalation decisionsStrong tierNative strengthMissed or late
RetriesCapped, then escalateOften unboundedOften unbounded
EvidenceCost per resolved conversationToken totalsToken totals

Frequently asked questions

Why are support agents more expensive than chatbots?

They run loops. Each step resends context, tool results add tokens, and retries repeat both. Cost scales with steps per resolution, not with the number of conversations alone.

Which steps can use cheap models?

Triage, routine replies, status checks and tool follow-ups. Escalation decisions, policy interpretation and conflict resolution benefit from stronger tiers.

How do I reduce context growth?

Return compact tool results with only needed fields, cache stable customer data for the conversation, and replace full history replay with a rolling summary.

When should an agent escalate?

After a second failure on the same tool or intent, on sensitive categories, or when confidence is low. Escalation is cheaper than repeated attempts and protects the customer experience.

How many steps should a conversation allow?

Enough to resolve routine cases and no more. Cap steps and tool calls, log where caps are hit, and treat frequent cap hits as a routing or prompt design signal.

Is there a moderation endpoint?

Not yet — it is coming soon. Apply your own input and output filters, and log safety events for review.

How does flat pricing affect agent economics?

Self-serve plans are flat monthly with no per-token charges, so loops affect latency and fair-use headroom rather than a variable bill. See the live pricing page for plans.

Can I build the first version for free?

Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger tiers for escalated cases.