Naveo

STEP 13 / 22

A5 TASK

YOUR PROMPT · 3 CASES

Orbit asks you to design the system prompt for the agent you built in step 11, now with explicit budgets communicated to the model. The code-side loop already has the limits; now the AGENT has to know them and respect them so it doesn't crash into them.

Three ceilings the agent must understand and report:

max_steps: 10 tool calls before having to give a final answer.
max_tokens_used: when it sees it's close to the token budget, it must condense.
max_cost_usd: when it sees it's close to the money budget, it must prioritize cheap tools.

The system prompt must teach the agent to:

Track which step it's on (the runtime injects it as context).
Warn the user when it's close to a ceiling, explaining what's going to happen.
Give a partial-but-useful answer when it hits a ceiling, instead of dying silently.

Where the user's goal goes, use {{input}}. The runtime injects each turn a <budget_status> section with steps_used, tokens_used, cost_so_far, and the remaining ceilings.

647 chars

use {{input}} where the input should go

RUBRIC · 3 CASES · 5 CRITERIA

"Get the Friday night shift roster and put together a summary of who's on…"

CASE 1

"Process the last 20 tickets, group them by bay, and return the 3 most ur…"

CASE 2

"List all crewmates with welding permits and check which ones have curren…"

CASE 3

GUEST MODE

You're viewing this lesson as a guest. To save your progress, earn XP, and keep your streak, sign in when you're ready to check.

Costs 1 heart

The agent knows its ceilings, doesn't crash into them

In step 11 you implemented ceilings on the runtime side: max_steps protects against loops, loop_detection against stuck, tool errors don't terminate the loop. That protects the system from model bugs.

But the other half is missing: the agent has to KNOW its ceilings and behave accordingly. If it only has runtime-side ceilings, when it hits them it stops abruptly, without warning, leaving the user with an empty answer.

The three classic ceilings

max_steps. how many tool calls before having to give a final answer. Typical: 10-15 for operational tasks, 30-50 for deep analyses.
max_tokens_used. how many tokens (in + out) the agent can consume cumulatively. Protects against context windows that balloon step by step.
max_cost_usd. how much money the agent can spend in this session. The absolute ceiling. No negotiation here: when it's gone, it's gone.

The system prompt teaches the agent to use them

Each turn the runtime injects a section like:

xml

The system prompt instructs the agent to:

Read that block before deciding the next tool call.
When at 80% of any ceiling, warn the user: "2 steps left. If you want me to go deep on X, that fits. If you want to cover everything, I'll have to cut sooner."
When it reaches a ceiling, give a useful partial answer: summarize what it accomplished, declare what's pending, offer to continue in another call.

What to avoid

Dying silently. The agent hits max_steps and returns null or a cryptic error. The user doesn't know what happened or what they have from the work done.
Returning only "budget exceeded". Throwing the error without a summary is mistreating the user. If you did 8 tool calls before cutting, those 8 have useful information. Summarize them.
Not reading the status. The agent decides blindly step by step, ignores the block, and is surprised when cut off. If it doesn't read, it doesn't act accordingly.

An agent well-trained in budgets looks considerate: warns, proposes, condenses, and respects the limits. An agent without budgets looks unstable: sometimes it finishes perfectly, sometimes it leaves the user hanging, and you can never tell which you'll get.

Your task

Write the agent's system prompt. Three ceilings to manage, a warning rule, a partial-answer rule. The judge evaluates five criteria on your prompt.