Job-Scope Routing and Catalog-Tool Ownership Boundary Automation: "It Doesn't Exist" Was the Wrong Answer
A queue that quietly accepts work it was never meant to do doesn't look broken — it looks helpful
Nothing about a task sitting in a general-purpose queue looks obviously wrong on its own. It has a title, a client, a deadline, an owner — it looks exactly like every other piece of legitimate work around it. The defect only becomes visible at the aggregate level: an entire category of work that was supposed to have moved, in full, to a dedicated system elsewhere keeps accumulating in the old queue anyway, and unless someone counts by type, it reads as normal volume rather than a standing routing leak.
How the underlying problem shows up before you fix it
A general-purpose queue keeps accepting a category of work that has since moved, in full, to a dedicated system elsewhere — nothing in the queue itself flags that the category no longer belongs there.
An agent's own permission boundary (which trades it can act for) and the shared tool catalog's actual contents are two separate sources of truth that nothing keeps reconciled — so "not permitted for you" and "doesn't exist at all" produce the identical negative answer.
A person asks an agent for something that agent genuinely cannot do, and the agent's own honest limitation gets reported upward as if it were a fact about the whole platform, rather than a fact about that one agent's scope.
A newly-mapped trade inherits a work queue by name — its stage appears on screen, tasks get assigned to it — without the underlying permission list being updated to match, so the queue silently returns nothing for that trade specifically.
A model-generated response occasionally embeds an invisible, zero-width character in the middle of an ordinary word, and nothing downstream strips it before that text is compared against anything else.
How a dated audit found both defects, and how the fix closes each one
The owner's rule, stated plainly during the same audit: no agent needs the general task queue to create social-media content — planning, review, approval, scheduling and publishing all live in one dedicated system, start to finish. Measured against the live queue at the time: fifty-two open tasks typed as "Social Media," twenty-two of them sitting specifically in the two stages meant for posting and scheduling, and twenty-three created by real people within the prior thirty days — active, ongoing misrouting, not old clutter nobody had cleaned up.
The fix removes those social-media stages from the trade-to-stage map entirely, so no trade's queue can surface a task filed under any of them again. The queue function itself was rewritten to additionally exclude any task explicitly typed as Social Media (catching cases that might not sit in one of the removed stages by name), and — just as importantly — to return a plain-language explanation of the rule alongside an empty or filtered result, so an agent surfacing "nothing in your queue today" also knows why, and can say so if asked rather than guessing. A house-wide training document stating the rule was added for every trade, and the training-lookup tool gained a fuzzy, word-overlap fallback after a real search using a paraphrased title came back empty — matching on shared significant words when an exact phrase isn't given, instead of failing silently on a request that clearly meant the same document.
A second, narrower fix in the same effort corrected a case where one trade's own queue access hadn't kept up with which trades actually had a stage mapped to them elsewhere in the task system — that trade's own agent told the owner the queue tool "didn't exist," when the real issue was a stale, hand-maintained permission list. Access now derives automatically from the same trade-to-stage mapping the queue itself already reads, closing the gap for good rather than patching the one trade found missing.
The more structural fix addresses a separate but related defect in how the platform's shared tool catalog is exposed to each agent. Before this fix, if an agent's own trade didn't have a specific tool listed, the response was unconditionally "this tool does not exist" — the exact same wording used for a tool that genuinely isn't in the catalog at all. Two real, separate incidents surfaced from this: staff members telling the owner directly that a specific tool didn't exist, in both cases when a colleague working a different trade had that exact tool available the entire time. The corrected dispatch now checks the shared catalog itself before answering: if the requested tool key exists there and is enabled, just scoped to different trades, the response names which trades DO have it and instructs the agent to route the request through the platform's own internal hand-off tool to a colleague in one of them — an honest "not mine, but real" instead of a false "doesn't exist." A smaller fix from the same pass strips invisible zero-width and byte-order-mark characters a model occasionally leaves in the middle of a word before any text-matching happens, closing a related, nearly invisible class of comparison failures found during the same audit.
What is actually built today
Posting/scheduling, publication, and social-media stages are no longer entries in the trade-to-stage map at all — no trade queue can surface a task filed under any of them.
The general queue tool also excludes any task explicitly typed as Social Media, and returns a plain-language explanation of the routing rule alongside an empty or filtered result.
Every trade with an actual stage mapped to it in the wider task system now automatically gets queue access — permission is derived from the same mapping, not maintained separately by hand.
A tool that exists in the shared catalog but isn't available to the requesting agent's own trade now returns an honest "exists, but isn't yours" answer naming which trades do have it, instead of a false "does not exist."
A house-wide training document states the routing rule for every trade, and the training-lookup tool now also matches by significant shared words when an exact title isn't given.
A shared text-cleanup function strips invisible zero-width and byte-order-mark characters a model can leave mid-word, closing a related text-matching defect found in the same audit.
A tool that exists, just not for you (illustrative framing of the actual measured finding)
An agent working the Art trade is asked for a tool that belongs to a different trade's kit. Before the fix, it reports the tool doesn't exist at all — an honest limitation reported as a platform gap that isn't real. After the fix, it reports the tool exists, names the trade that has it, and routes the request to a colleague there through the platform's own hand-off tool, instead of leaving the requester with a wrong impression of what the system can do.
What changes operationally
Centriu Axis's AI staff agents can no longer misreport a real, working tool as nonexistent just because it isn't theirs, and social-media work can no longer sit inside a queue meant for everything else — closing both a routing leak (fifty-two tasks, nearly half created in the prior month) and a trust leak (two independent false "doesn't exist" reports that reached the owner directly).
When this is not the right fit
This fix governs which trade's queue a task appears in and how an agent describes a tool outside its own trade — it does not change what any specific tool actually does, and it does not grant a trade access to a tool that is deliberately scoped to a different one.
A flat "it doesn't exist" vs. an honest "it's not mine"
A tool-dispatch layer that answers every out-of-scope request with the same blunt negative used for something genuinely absent is, functionally, indistinguishable from a router that has no idea what exists elsewhere in the building. Centriu's fix treats "not available to you" and "doesn't exist anywhere" as two different facts worth two different answers — the same distinction a competent front-desk transfer makes between "that service isn't offered here" and "that's a different department, let me connect you."
Related systems
Main system: Centriu Axis. Complementary when relevant: Centriu Atlas.
What it does NOT do
- Does not grant a trade access to a tool that is deliberately scoped to a different trade — it corrects only the wording used to explain that boundary.
- Does not change how a genuinely nonexistent tool key is reported — that continues to return a real not-found response, unchanged.
- Does not automatically move any existing Social Media task into the dedicated system — it stops that category from appearing in the general trade queue going forward; anything already filed under the old stages still needs to be relocated at the source.
- Does not review or rewrite training content itself — the fuzzy search only changes how an existing training document is found, not what it says.
- Does not strip zero-width characters from content already stored before this fix — the cleanup applies to text as it is processed going forward.
- Does not replace the platform's internal hand-off tool — it points an agent toward using that existing tool rather than introducing a new one.
Security and governance
Every queue and catalog check resolves the requesting agent's own trade and organization before returning anything, and a hand-off between trades stays within the same organization's own agents. Any business data referenced remains subject to Brazil's LGPD (Law No. 13,709/2018). Full detail on access control lives at /governanca and /iso.
Pricing and contracting
Included at no extra cost with any Centriu contract. Values and terms come from the official pricing table at /precos (Centriu's central source — never restated here).
Frequently asked questions
Why were social-media tasks appearing in the general task queue at all?
The trade-to-stage map had, at one point, pointed the posting/scheduling and publication stages at specific trades — a routing choice later reversed once the rule that all social-media work lives in one dedicated system was made explicit.
How did an agent end up telling someone a real tool didn't exist?
The tool-dispatch logic answered any request for a tool outside the requesting agent's own trade with the same flat "does not exist" wording used for a tool genuinely absent from the catalog — it never checked whether the tool existed elsewhere first.
Does the fix let every agent use every tool now?
No — trade boundaries are unchanged. An agent still can't use a tool outside its own trade; it now says so honestly and points to the colleague who can, instead of denying the tool exists.
What was the zero-width character issue?
A model occasionally left an invisible, zero-width character in the middle of an ordinary word in its own output; a shared text-cleanup step now strips it before that text is compared against anything else.
How was this verified?
Directly against real production data: one trade with no mapped stage now returns the explanatory message instead of a stale queue view, a second trade's queue was checked and confirmed to contain zero Social Media items, and a real paraphrased training search correctly found its matching document.
What does Centriu Axis cost?
It is sold by subscription with a published starting price — exact current values are on the central pricing page.
See how Centriu Axis keeps AI agents inside their own lane
Reach our commercial team directly, or leave your details below — we'll follow up with guidance for your case.