Your first conversation
You already know how to give an agent a task. Let's follow what happens between your message and its answer. We'll start with your own project, then inspect a recorded conversation with Navlog. You can explore the recording without installing or running the example.
The video follows two read-only requests in Navlog, a B4 example. Watch the file read, the answer, and the follow-up. We will inspect this same conversation below.
Start with a request you can checkCopy link to section: Start with a request you can check
Choose your connected Codex or Claude Code agent and a project folder, then ask:
Explain what this project does and where its main entry point is. Do not change any files.
The first message starts the conversation. The agent may read files before it answers. Follow that activity as it arrives, then check the answer against the project. A useful first result names the entry point and explains how it fits into the application.
Understand a turnCopy link to section: Understand a turn
A turn begins with your request. It includes the agent's work toward that request and ends when the runtime reports an outcome. One turn can contain several tool calls. A tool call might read a file, run a command, or ask a server for information.
A tool result tells us what came back from that action. The final response explains the outcome to you. Read both when you need to understand how the agent reached its answer.
Respond when the agent needs youCopy link to section: Respond when the agent needs you
Some actions require a decision. Read the requested action before responding. The available choices depend on the runtime and its configured permissions. A waiting turn needs your response before it can continue.
An approval allows the requested action; it does not prove that the action succeeded. Follow the subsequent result.
Ask a follow-upCopy link to section: Ask a follow-up
After the turn finishes, ask a narrower question in the same conversation:
Which file would I read next to understand how that entry point works?
You are continuing a conversation with the same runtime. Starting a conversation with another provider does not transfer hidden context between providers.
Stop or recoverCopy link to section: Stop or recover
Use the conversation's stop control when you want to interrupt active work. Wait for the reported outcome. If the runtime disconnects before confirming the result, the app may show an unknown outcome. Do not assume the request failed or resend it blindly. Follow the recovery action shown by the app.
Follow the same idea in NavlogCopy link to section: Follow the same idea in Navlog
Navlog is a B4 example that prepares a flight plan using tools and source material. A request leads to lookups, results, and a response, just as a coding task leads to file reads and an explanation. Its weather is time-dependent, and its local flight-plan recording is not a transmission to a filing service.
Let's look at a smaller request first. In this recording, we ask Navlog to read its aircraft profile. It calls readDoc, checks for remembered preferences, and answers. We then ask a follow-up in the same conversation.
This is a recorded conversation. Read the transcript or return to the lesson.
Read the tool's arguments and resultCopy link to section: Read the tool's arguments and result
Inspect the first file read. The inspector selects the readDoc call at the point where its result arrived.
Choose Arguments. The path value is aircraft/c172n.md: the file the agent asked the tool to read. Now choose Result. This is what the tool returned. We can compare the requested path with the returned contents before reading the agent's explanation.
The Tree view makes the returned data easier to explore. Choose Show full value to read the complete file. This particular result arrived as a JSON string; the inspector labels the parsed view. JSON shows the original recorded value.
There are two useful questions to ask here: did the tool receive the input we expected, and does its result support the answer?
The request names the file directly. Its profile describes a Cessna 172N with 50 US gal of usable fuel and cruise power at 2400 RPM. Compare those entries with the answer.
Follow the next turnCopy link to section: Follow the next turn
Inspect the follow-up turn. We asked about reserve and alternate fuel, and asked the agent to read the profile again. The same tool reads the same file, but this is a separate call with its own arguments and result.
Choose Previous event in the player. The result becomes unavailable because we moved to a point before it arrived. Choose Next event to bring it back. The recording lets us check what the app had observed at each step.
Inspect the completed conversation. Read the final response alongside the file. The profile includes a reserve entry. Look for whether it also contains the alternate-fuel information the follow-up requested. A useful answer should distinguish what the source says from what it does not provide.
You can select another message or tool to inspect it, or use Play to follow the recorded timing. The speed control changes playback speed; the elapsed time still describes the recording. On a narrow screen, use Conversation and Details to switch between the two panes.
Connect the example to its sourceCopy link to section: Connect the example to its source
If you build B4 agents, choose Source to read the example's readDoc tool and aircraft profile at the recorded revision. The source explains how the tool works. The replay shows what the app observed when it ran.
This recording uses B4 beta 0.1.0. It reads example files; it does not fetch weather or file a flight plan.
You can run the B4 example when you want to explore that path. To keep your conversations organized, continue with manage your work.