LlamaIndex Agents and Workflows: Keeping Retrieval and Actions Separate

Evidence check: 25 October 2026. Product details reflect the referenced documentation observed at this cutoff; exact plans, prices, regional eligibility and unverified release dates are not claimed. This is an editorial guide, not a hands-on test.

Start with grounded data

A knowledge-heavy agent needs document identities, timestamps, permissions and retrieval evaluation before it needs an elaborate hierarchy of workers. LlamaIndex’s framework documentation covers agent building and associated tools, while its workflow documentation describes composing application steps. Agents can ask tools for information, but a tool response still needs source provenance and bounded authority.

A version-aware research task

Suppose a reader asks which revision of a safety policy applied to a product shipment. Build an index that preserves effective dates and market coverage, then use an agent with a policy-retrieval tool to select relevant material. If the supporting document is absent or historical versions conflict, generate a review request instead of guessing. The agent may draft a report, but should not change an official policy record. Separating these outcomes makes the editorial record auditable.

Workflow control

Represent retrieval, extraction, claim reconciliation and human approval as explicit steps where possible. A paused process should store a stable record ID and a draft version, not only a conversation transcript. If the task requires output to an external database, validate the final payload and recheck permissions immediately before the write. An interrupted operation must be reconciled with external state before replay; prompts cannot provide transactional safety.

Test retrieval and planning independently

Run source-recall tests without generation to confirm the right passages are available. Test planning with mock tools to identify unsupported calls. Then run end-to-end tests that detect stale citations, cross-tenant leakage, missing source IDs and duplicate drafts. Hold out a portion of test questions to reduce tuning to a memorized benchmark. Record framework version and retrieval configuration because index updates can change results even when the underlying model is unchanged.

Trade-offs

LlamaIndex can be appropriate when retrieval is the center of the agent’s work. Haystack offers another pipeline-centered architecture; LangGraph emphasizes explicit state transitions; direct retrieval and template rendering may be enough for simple tasks. Check deployment, data residency and third-party integrations separately. No independent performance result or provider price has been established by the documentation alone.

Editor’s audit example

For each extracted claim, save the original document identifier, retrieval span, effective date, originating market and confidence state. Then ask a separate reviewer to confirm whether the cited passage actually supports the exact wording. This audit is distinct from a model-generated source list. If the agent mentions a policy that is no longer current, the system should link to the superseding document rather than silently rewriting the history. Such practices are framework-independent and should persist if the team changes orchestration libraries.