Problems

Build a Thought -> Action -> Observation loop that terminates on a Final Answer.

Parse agent output and route tool calls to the right handler with a fallback for unknown tools.

Wrap tool calls in try/except so that runtime errors become observations and the loop continues.

Track step count and return a fallback result when the agent exceeds its allowed step budget.

When the scratchpad exceeds a token limit, summarize the older portion while preserving the latest action and observation.

Execute a list of ordered plan steps by dispatching tools in sequence, accumulating observations, and returning a final answer.

Design an approval gate that classifies actions by risk level, creates sanitized approval requests, executes only on explicit approval, and produces an auditable trace.

Select a model from cost, latency, and capability metadata: reject candidates exceeding hard budgets, choose the least expensive model that satisfies quality requirements, and return selection reasoning with estimated spend.

Route subtasks to specialised workers by capability, respect task dependencies, isolate worker failures, and enforce a global step budget.

Integrate tool routing, memory retrieval, RAG citation, prompt-injection checks, bounded execution, and safe fallback into a single guarded agent loop.