Scoping a Custom AI Agent: What Actually Drives Effort in 2026

Why AI agent projects vary so much in scope

Two AI agents that sound similar — "triage my inbox" — can differ by 10× in engineering effort. This post explains the factors that actually drive scope, so you can scope better and communicate more precisely with whoever builds your agent.

Factor 1: Integration breadth

Every system the agent touches multiplies complexity:

• 1 integration (e.g., just Gmail): Compact. One API, one authentication flow, one data model.
• 3 integrations (e.g., Gmail + Notion + Slack): Moderate. Auth × 3, data mapping × 3, error handling × 3.
• 5+ integrations: Extended. Cross-system state, conflict resolution, and complex error recovery become central concerns.

Factor 2: Reasoning depth

• Classification/routing: "Is this email about billing, support, or sales?" — Compact. The LLM makes one decision.
• Multi-step with data lookup: "Check the customer's order history, then decide whether to issue a refund" — Moderate.
• Multi-step with ambiguity: "Understand what the customer actually needs, even if they didn't say it clearly" — Extended.

Factor 3: Observability requirements

A basic agent logs what it did. A production agent needs:

• Run traces: Every decision, with reasoning
• Human review gates: For actions above a confidence threshold
• Cost attribution: Per-run, per-model cost tracking
• Anomaly detection: Alerting when behavior changes

Observability can easily be 30% of the total scope.

Factor 4: Infrastructure ownership

• Your infra, your maintenance: More control, more responsibility. Khtain default.
• Managed platform: Less control, less responsibility, monthly fee.

Factor 5: Edge-case rigour

What happens when:

• The LLM returns malformed JSON?
• The API is down for 10 minutes?
• A customer sends a 40-page PDF attachment?
• Two triggers fire simultaneously?

Each edge case you handle doubles the reliability — and adds scope.

Real examples (anonymized, approximate)

• Notion-to-Slack triage agent: 1 integration chain, deterministic routing with LLM fallback, runs on founder's VPS. Compact.
• Multi-source RAG system: 4 data sources, custom chunking, self-hosted on Hetzner. Moderate.
• Multi-chain Discord bot: 2 blockchains, token-gating logic, stateful interactions. Moderate-extended.

The scoping call

The best way to get a real scope estimate is to talk through your specific use case. Khtain's scoping calls are free, 30 minutes, and you'll leave with a concrete understanding of what your agent needs — even if we don't work together. Book here.