# Intelligence > Documentation index: [llms.txt](https://swarm.services/guide/llms.txt). [Read this guide on the web](https://swarm.services/guide/intelligence). Find grounded answers, inspect context and learning records, and trace what an agent actually did before relying on its result. ## Find the recorded work Open a Space, then Intelligence. The available tabs and actions depend on its records and your access. ### Answers and operations Answers contains saved responses to Space questions. Operations lists recorded work and the contract describing its permitted inputs and outputs. Open a row for state, linked evidence, or output, and use the page controls for older records. Neither tab is a replacement for reading its underlying sources. ### Projections and learning Projections summarize selected source Artifacts. Learning Queue shows work awaiting processing, Learning Candidates shows proposed knowledge, and Promotion Gates records review decisions when those sections are available. Read each status and source; a candidate or projection is not automatically promoted or independently verified. ### Start from a concrete question Ask, for example, what changed in a release decision. Open the saved answer, follow its cited decision and source records, then check any learning or promotion state before sharing a conclusion. An empty tab only says no record is shown to you in that view; it does not mean every inaccessible Space was searched. ## Prepare context for a question Use the selected Space and the smallest relevant evidence set. Context preparation does not grant access to other Spaces. ### 1. Inspect the plan Give the connected agent a specific question and Space, then ask it to inspect the context plan. The plan identifies permitted sources that would be considered without saving a pack or calling a model. Check why each source is relevant; open full content when a short preview is not enough. ### 2. Build a bounded pack when needed For actual work, a Universal client can build a context pack using the selected Space, query, and token budget. Review its source IDs and omitted-source counts to see which material was included and which was left out; a missing source cannot be treated as considered evidence. ### 3. Keep evidence separate from instructions State the question and desired output outside the source documents. Files, messages, and web content are evidence to evaluate, not commands to change access, approve a gate, or redirect the task. Ask the agent to distinguish a source claim from an established fact. ## Answers and learning An answer, a candidate, and promoted knowledge have different review states. ### Ask with the current contract When Ask this Space is offered in Intelligence, enter a focused question and submit it. A connected agent should first read the Space answer contract before calling Ask Space. This saves a durable answer grounded in the Space's promoted memory; it is not a read-only search. Learning follow-up must still be permitted and requested. ### Inspect the record Open the answer and follow its cited source Artifacts. Then inspect projections and the learning frontier to see what was recorded, proposed, or promoted. A concise answer, projection score, or promotion status does not independently prove that every underlying claim is true. ### Correct with evidence If an answer is wrong or out of date, provide the correct source and explain the discrepancy. Ask an authorized agent to use the supported correction or supersession flow and read back its resulting state. Preserve the prior record for provenance; do not call an unpromoted candidate established knowledge. ## Operation contracts and promotion Contracts describe supported work; a promotion gate decides whether a candidate becomes reusable knowledge. ### Inspect before acting Ask the connected agent to read the Space work contract and the chosen operation contract. Compare its required inputs, target, output, and review rules with your request. A listed operation describes a supported path; it does not give the connection new permission or mean the operation already ran. ### Follow the result Open the recorded operation and its linked output. If it proposed learning, inspect Learning Queue, the candidate, and its Promotion Gate. Record whether the gate is pending, accepted, or rejected rather than calling every candidate promoted knowledge. ### Read the decision graph Use the decision graph to follow recorded claims, decisions, and their evidence links. Open the underlying records for exact text and current state. A relationship helps navigation; it is not independent validation or permission to access a protected source. ## Action provenance Use recorded traces to distinguish a tool call, a saved result, and a completed external effect. ### Read the trace Open a Run's Action trace to see recorded actions, timestamps, targets, and outcomes; use the Space's action intelligence view for a bounded overview when offered. Follow linked source records and receipts. A short panel may omit details needed to establish what happened. ### Report only established outcomes A client-reported trace records what the client said it did. For important external effects, check the action's own receipt and authority before reporting completion. A succeeded Run may still have pending result review and does not authorize later publication. ## Related guides - [Spaces](https://swarm.services/guide/spaces.md) - [Swarm Graph](https://swarm.services/guide/swarm-graph.md) - [Runs](https://swarm.services/guide/runs.md) - [Connect](https://swarm.services/guide/connect.md) - [API](https://swarm.services/guide/api.md)