Track 04 · Multi-agent & integrationLesson 10 of 11
Live web search as a tool
Bring in market and regulatory facts, with sources and freshness checked.
- Lesson library · reference
- 35 min read
- 60 min hands-on
- Pro
What you'll be able to do
01- Decide when an agent needs live web retrieval, when it needs a licensed data feed instead, and when it must not go outside the estate at all.
- Choose between provider-native search tools and third-party search APIs, using control, citations, data handling and cost.
- Implement a governed search wrapper: redact internal identifiers from queries, enforce domain allow-lists, deduplicate, filter by recency and format citations.
- Defend against indirect prompt injection from fetched pages, and explain the residual risk.
- Design caching, cost controls and audit logging for search, and answer the compliance questions (robots, terms of service, copyright, query leakage).
Before you start
- Answers with receipts: grounding and citations: answer contracts and citation verification. Web results are evidence like any other.
- Retrieval that plans, checks and retries: retrieval as a tool, and why web fallback is a governance decision rather than a default.
- Automating systems with no API: page content as untrusted input.
- Designing tools a model can use well: designing the search tool's parameters and results.
The short version
Some questions cannot be answered from the customer's own data: "Is Rotterdam congested this week?", "Has this carrier announced a surcharge?", "Is this consignee on a sanctions list?", "What does the new EU customs rule require?". Giving an agent a web search tool answers them, and opens three new risks: confidential information leaving the estate inside queries, low-quality or stale sources reaching an answer, and web pages that contain instructions aimed at your agent. A production design treats search as a governed integration. It redacts internal identifiers before a query leaves, restricts results to domains the customer has agreed to trust, prefers recent and authoritative sources, deduplicates, keeps fetched text in a delimited untrusted block, cites every claim with a URL and date, logs what was sent and received, caches sensibly, caps spend, and uses a licensed feed instead of search wherever the answer matters legally or financially.
Key takeaways:
- The query is data leaving the estate. Redact before you search.
- Allow-lists beat block-lists for enterprise answers; widen them deliberately.
- Fresh is not the same as authoritative. Score sources on both.
- Fetched pages are untrusted input. Never let them steer tools or writes.
- For sanctions, tariffs and regulatory status, buy or subscribe to the authoritative feed; use search to find what to read, not as the record.
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI