The agent picks Claude Sonnet 4.6 from the model selector. It calls the quote-drafting tool, then the AR-lookup tool, then the email-send tool, all in one tool-calling loop that this module runs end to end. The system prompt is cached on the second turn, so token costs drop. A 503 from the API mid-loop is retried with exponential backoff that honors Retry-After. The Anthropic request-id is captured in the audit log, so a support ticket points at the exact upstream call. The CFO gets the email; the audit trail shows every tool call, every model, every retry.

When the one person who knows everything is on leave

Every team has a human search engine: the colleague who knows that a customer gets net-45 because of a verbal deal, that one building uses a different SKU prefix, and that you never trust cost fields imported before March. None of it is written down anywhere. When that person is in a meeting, work queues behind them; when they quit, a slice of the company's operating memory walks out the door. The real cost is the daily drip of a hundred quick-question interruptions nobody ever logs.

What the module does. An admin builds an Odoo AI Agent, sets its model to Claude, attaches the rate sheets and the how-we-actually-do-it notes as the agent's source documents, and wires the live-data tools onto its topics. Colleagues ask the agent instead of pinging the one person. The module runs the full multi-turn tool loop inside Odoo: Claude calls a configured tool, reads the result, calls another, and answers, bounded by the iteration caps and limited strictly to the tools the admin wired up.

Turn the go-find-X, then-Y, then-draft-Z chain into one ask

A customer asks you to re-send last spring's unpaid invoice and confirm the delivery address on file. Sounds like thirty seconds. Instead someone opens Sales, finds the order, jumps to Accounting, filters invoices, opens three to find the unpaid one, opens the contact form for the address, then drafts the reply. Each hop is trivial; done forty times a day across a support or receivables desk, it is the whole day, and it never shows up on any process map.

What the module does. The module lets an Odoo AI Agent run that chain on Claude instead of a person. The admin configures the agent's topics with specific tools, and the module runs the loop end to end: Claude calls the first tool, reads the result, decides the next, chains through them, and writes the final reply, up to the configured caps. Per-agent reasoning effort and adaptive thinking let it work out which invoice the request maps to on harder cases without you switching models.

The 2 a.m. dropout that never showed a red banner

An agent that drafts dunning notes or pre-classifies bills does its best work overnight, exactly when nobody is watching. One night the endpoint returns a 529 or a 503, the call falls over, the run produces nothing, and there is no error on screen. By morning the agent looks healthy. The gap surfaces at month end, and in the post-mortem all you can hand support is that the AI did not work last Tuesday, which dies at we cannot reproduce it.

What the module does. The module sits between the agent and the Anthropic Messages API and turns one-shot failures into recoverable ones. Every Claude call retries automatically on 429, 500, 502, 503, 504, and 529, backing off exponentially with jitter and honoring any Retry-After header, so a transient overload is absorbed rather than fatal. When a call truly cannot recover, it reads the Anthropic request-id from the response, writes it to the log, and surfaces it in the user-facing error with the HTTP status, so a ticket points to one exact failed call.

Stop paying full price for instructions Claude already read

Your best agents carry the heaviest system prompts: the refund policy, escalation rules, tone guide, a page of never-do-this. That block is long, stable for weeks, and genuinely keeps the agent on-policy. The part nobody demos: every message in the conversation pays for the whole block again. A four-step task re-reads the entire policy preamble four times, and the instinctive fix, trimming the prompt, is exactly backwards because it weakens the guardrails you wrote it for.

What the module does. Switch on Cache system prompt for that agent. The module then attaches an ephemeral cache_control marker to the agent's flattened system prompt on every call, so Anthropic writes the stable prefix to its prompt cache once and eligible later requests reuse it instead of reprocessing it. Because the multi-turn loop re-sends the conversation each turn, the savings land where the pain is: later turns of a task and quick follow-ups ride the cached instructions, with no model change and no trimming.

Show me where you got that, with a real answer

Your team already had an AI assistant and quietly stopped trusting it, not because it was often wrong but because it was wrong invisibly. Someone asks for a customer's agreed payment terms and gets a clean, confident paragraph with a number that might be from the signed terms or might be the model averaging what terms usually look like. There is no way to tell from the answer, so careful staff re-check the record anyway and the tool saves nothing, while juniors act on it and twice a quarter quote a wrong figure.

What the module does. The agent answers about company-specific data only by calling the tools an admin configured, and the module runs that loop against your live records, so a figure about a customer's terms comes from a tool reading that record now, not from training. Where source documents are attached and an embeddings key is set, retrieval tags each chunk with its source name and attachment id and surfaces those references with the answer, so where-did-you-get-that has a literal answer. When a call fails, the request-id makes a disputed answer a traceable event.