Explore more topics
Developer field guides

Build, evaluate and observe AI agents: developer guides

Follow NexusTrade's agent loop, tool dispatch, evaluations and event architecture through worked examples. Choose the implementation problem you need to solve.

As of 2026-10-10

Jump to a section

What are you building?

Three implementation tasks: build an agent loop and tools, evaluate outputs, and observe model and trading events. Framework selection is a separate decision.
Pick the part you need. You can inspect an existing system without rebuilding its agent framework.

Build a ReAct agent loop

Follow a decision through tool execution, observation and the next decision. Inspect where the loop stops.

Open page

Build a tool-call framework

Validate arguments, dispatch the right handler and return useful errors. Run the local worked example.

Open page

Evaluate LLMs

Freeze representative tasks, separate deterministic checks from grading, and inspect failures alongside cost and latency.

Open page

Inspect LLM traces

Follow a run through nested spans and tool calls. Learn what to record and how to keep customer traces private.

Open page

Diagnose trading strategy events

Read the event history to explain a strategy's behavior, separating signals, orders and fills.

Open page

Choose LangChain or a custom framework

Compare implementations of the same local task and inspect the boundaries you would maintain.

Open page

Start with the failure or output you need to inspect

Show the table · 6 rows, 2 columns
TaskConcrete starting artifact
Build a loopA local two-decision, one-tool exposure calculation returning 80.00%.
Dispatch toolsSPY 20% preview plus INVALID_ARGUMENTS and UNKNOWN_TOOL receipts.
Read model tracesA local summary with 140ms step time and two unknown-cost steps.
Diagnose tradingTwo fixture records: one BuySignal and one OrderRejected, with zero Order records.
Compare frameworksThe same handler and correlated receipts in a custom dispatcher and LangChain adapter.
Evaluate outputsFreeze requests and inspect each expected/actual judgment and failure, rather than only an average score.

These local outputs are downloadable teaching fixtures. Choose the implementation task in the guide list to run its exact inputs and inspect its result before connecting a model or account.

Here to connect your app instead?

These guides explain implementation. For account-scoped SDK or MCP access to NexusTrade, start with the developer setup page.

Set up NexusTrade SDK and MCP access ↗

Discussion

Sign in or create a free account to join the discussion.

No comments yet.