A task-completion agent for library operations

Give library staff their afternoons back.

Stacks resolves room-booking conflicts, routes ambiguous interlibrary-loan requests, and chases overdue items, automatically when it's safe, and with a staff member in the loop whenever it isn't.

Your systems already flag the hard cases. Stacks is what works the queue they flag them into.

No setup for staff. Real Cognito sign-in, real AWS infrastructure underneath, built on Amazon Bedrock AgentCore.

Live case walkthrough

Case received
Evaluating policy
Memory recall
Risk assessment
Safe to automate

Resolved automatically

The problem

Every library system has an exception queue.

Room-booking software, interlibrary-loan platforms, and ILS overdue modules already automate the routine case well. What they hand back is the case a rule could not fire on: an incomplete citation, a genuine policy conflict, a patron history that changes what the right action is.

ROUTINE, ALREADY AUTOMATED~70%

EXCEPTION QUEUE, LEFT TO STAFF~30%

Illustrative, not universal: OCLC's own Penn State member story documents roughly 30% of ILL borrowing requests staying manual even with automation in place. Figures vary by library and workflow.

The difference

Stacks starts where the existing system stops.

Without Stacks

Request
Rule
No match
Review queue
Staff

With Stacks

Request
Context
Policy
Specialist
Safety
Resolution

Three real workflows

Not features. Cases Stacks actually completes.

It doesn't just answer

Library chatbots inform. Stacks completes the case.

Library AI assistants such as USF LINK and SJSU KingbotGPT are genuinely useful, information-oriented systems: they help patrons find services, collections, and spaces. Neither is built to take a transactional action on a case. That is a different job, and it is the one Stacks does.

Library chatbot

QuestionAnswer

Stacks

CaseContextDecisionActionOutcome

Safety

The model doesn't choose its own safety tier.

A plain, deterministic classifier, ordinary code with no model call in it, decides GREEN, YELLOW, or RED for every case. The agent cannot argue with it, and it cannot be prompted around.

Case + sensitivity signalsCode-governed gate

Auto-executedAwaiting confirmationRequires review

Policy

Stacks doesn't invent the rule. It cites the rule.

Every resolution names the exact policy clause it applied, verbatim, the same clause a staff member sees in the real approval screen. Not a generic explanation after the fact: the rationale is required before the tool will commit anything.

Policy applied

RBP-1: A recurring, library-run program outranks a one-off renter or walk-in booking for the same slot.

Shown exactly as it appears on a real case in the Stacks console.

The system, end to end

Six real components, working as one

Stacks Agent

The orchestration layer. Reads each case, calls the right tool, and never invents a resolution the tool itself didn't return.

ILL Disambiguation Specialist

A dedicated agent that narrows an ambiguous interlibrary-loan request to a single confident candidate. The main agent cannot claim convergence unless it actually happened. That claim is verified in code, not asserted in a prompt.

Amazon Bedrock, agents-as-tools

Overdue Escalation Sequencer

Runs the nightly reminder ladder as a durable, session-backed process: resumable, never double-sent, never stuck mid-tier.

Bedrock AgentCore Runtime + EventBridge

Bedrock Guardrails

A real, live guardrail policy screens every outbound notification for sensitive content before it reaches a patron.

Amazon Bedrock Guardrails

AgentCore Memory

Recalls a patron or requester's real history, prior hardship flags, prior substitutions, so staff see context, not a cold case.

Amazon Bedrock AgentCore Memory

HITL Approval Gate

A code-governed classifier, never the model itself, decides GREEN, YELLOW, or RED, and RED cannot resolve without a qualified human.

Deterministic, not a model decision

Audit

Every case leaves a record, not a mystery.

A staff member can always answer "why did Stacks do that" from structured evidence, policy matched, context used, specialist consulted, safety classified, action taken, never from exposed model reasoning.