Cohere Tool-Using Agent Loop
Overview
Let the model propose typed tool calls while application code validates, authorizes, executes, correlates, and limits every side effect.
Prerequisites
- The target repository, runtime, environment, and accountable owner
- An approved Cohere team and key for any live verification
- Current quality, security, privacy, capacity, and change-control requirements
Tool Discipline
Use Read, Glob, and Grep to inspect code, configuration, and evidence. Use WebFetch only for current Cohere primary documentation. Use Write or Edit only when the user requested implementation and the exact target files are known; never write credentials or customer content.
Current Contract
- Define each v2 tool as a function with JSON Schema parameters.
- Read proposed calls from the response message and correlate results with
tool_call_id. - Append the assistant tool-call message and tool results to the application-managed
messageshistory. - Command A+ was the current high-capability tool-use candidate on the review date; resolve availability before pinning it.
Authentication
Use an environment-specific key injected from an approved secret manager. Never print, persist, commit, or place CO_API_KEY in an example. Confirm access with the least costly bounded operation appropriate to the task, and treat key creation, rotation, revocation, role changes, and production-capacity requests as owner-approved actions.
Instructions
- Define a minimal tool allowlist, strict schemas, timeouts, and side-effect classifications.
- Resolve a live tool-capable model and submit the user message plus tool definitions.
- Validate tool name, arguments, tenant scope, and approval state before dispatch.
- Execute read-only calls automatically only if policy allows; pause for consequential writes.
- Append correlated tool results and continue until content is returned or the iteration budget is exhausted.
- Test unknown tools, invalid arguments, duplicate calls, timeouts, denial, and loop exhaustion.
Approval Boundaries
Do not expose or rotate keys, change Cohere Team roles, accept commercial terms, enable sensitive production data, increase spend or capacity, switch production models, send a support bundle, or execute model-proposed side effects without the accountable owner's approval. Keep diagnosis read-only unless implementation was requested.
Output
Return the resolved API and model contract, files or settings inspected, evidence collected, validation result, remaining risk, owner, and rollback or next action. Redact keys, authorization headers, prompts, retrieved documents, embeddings, customer identifiers, and unrestricted environment output.
Error Handling
| Condition | Response | |---|---| | Unknown tool | Return a structured denial result; never dispatch by reflection. | | Invalid arguments | Reject before execution and preserve validation evidence. | | Missing correlation | Stop because tool results cannot be safely attached. | | Loop exhausted | Return partial state and request direction without another model call. |
Examples
Use this compact handoff shape to keep the selected scope, validation evidence, and operational result reviewable.
Input:
goal=lookup-order; tools=read-only-order-status; max-iterations=4
Expected handoff:
iterations=2; tool-calls=1; side-effects=none; final=validated