The Depot — Mini-Agent Playbook
A governed toolkit marketplace for the workshop floor. Today it ships prompt cards. The next step is turning the highest-value cards into governed RAG agents that return compliance-grade answers.
Executive summary
The Depot is a single place where coders and non-coders browse proven mini AI agents and skills and put them to work. Today it ships as prompt cards: copy a prompt, paste into an assistant, run. That is a strong start and a weak ceiling.
The next step is the one already spotted: turn the highest-value cards into governed RAG agents, real tools that ingest source documents, retrieve with citation, and return compliance-grade answers. The LTA Code of Practice agent is the flagship example.
This playbook answers three questions:
1. What else can we build — quickly, robustly, consistently, at enterprise grade.
2. How do we govern it — a framework that is real but not overdone, and keeps dev + deploy easy.
3. What happens after launch — the blindspots we are missing today, and how to reduce adoption friction from senior management to the floor.
One scaffold, many agents
The LTA agent, a tender-compliance agent, and every future regulator agent share the same architecture.
Build that scaffold once. Each new agent then reduces to two steps: point it at the source documents (whitelist), and write a golden test set (known questions with expected clause citations).
The current prompt-card model cannot deliver this. A pasted prompt has no audit trail, no citation guarantee, no version control. The governed agent does.
Governed by risk, not one-size-fits-all
The tension between "proper AI governance" and "not overdoing it" is solved by risk-tiering. You do not govern a read-only document-lookup agent the same way you govern an agent that files corrective actions. Pay for the control you need.
| Tier | What it is | Example | Governance required | Time to ship |
|---|---|---|---|---|
| T0 | Skill / prompt card. No data, no actions. | Current Depot cards (Shift Handover, Fault Code Translator) | Human review of the prompt + a "no sensitive data" rule | Hours |
| T1 | Read-only RAG over whitelisted docs. No PII, no writes. | LTA Code of Practice agent, Tender Compliance | Source whitelist + citation-forced + golden test set + human sign-off for decisions | Days |
| T2 | Acting agent. Writes, external calls, PII. | Safety Incident → Corrective Action (escalation path) | Everything in T1 + data classification + access control + audit log + approval gates + rollback | Weeks |
The pipeline: Plan → Build → Test → Deploy
1. Plan — a one-page spec (goal, users, source docs, risk tier, success metric, owner). A 15-minute fill-in, not a project.
2. Build — clone the standard RAG scaffold. No new architecture per agent.
3. Test — golden set (≥20 known Q/A pairs with expected clause citations), hallucination spot-check, PII scan, currency check (source version + last-updated date).
4. Deploy — register in The Depot with a tier badge, version, "last updated," owner, and audit trail. Rollback = revert to previous version.
LTA Code of Practice agent
The reference implementation. This is the agent made concrete.
Sources: the LTA codes that govern train maintenance, led by the Code of Practice for Railway Protection (2024 Edition), plus the related maintenance and design codes. Four PDFs in the initial corpus.
Tier: T1 — read-only RAG over public regulatory PDFs. No PII, no writes. Fast-track.
Users: project teams (compliance before submission), maintenance teams (implementation steps), compliance officers (audit).
Why it matters: it turns a 3.7MB PDF that nobody reads into a working compliance partner. Fewer audit findings, faster sign-off, consistent interpretation across teams. The regulatory basis sits in the Railway Protection & Road Structure Safety Zones framework, underpinned by the Rapid Transit Systems Act / Regulations.
Two more quick-win agents
Two ideas that reuse the LTA scaffold, are quick to build, and demonstrate the framework's range.
Idea 1 — Tender / Bid Compliance Agent (Tier 1)
What: ingest the RFP/tender spec + our draft submission → mandatory-requirement checklist → compliance matrix → gap flags → suggested evidence for each.
Why: direct commercial ROI — fewer disqualifications, more wins. Same RAG scaffold as LTA, pointed at a different corpus.
Users: bid teams, project managers, compliance. Tier: T1, fast-track.
Idea 2 — Safety Incident → Corrective Action Agent (Tier 1 → 2)
What: raw incident note → structured report → 5-Why root cause → suggested corrective actions → auto-checked against the Work Instruction library and the regulatory codes → escalate if safety-critical.
Why: it closes the loop — it connects floor data to the compliance corpus. This is the agent that turns incidents into compliant, auditable action instead of a report that sits in a drawer.
Users: floor technicians, supervisors, safety officers. Tier: T1 for the analysis path; the escalation/write path is T2 (approval gate before any corrective action is filed).
How the scaffold compounds
Once the scaffold exists, the marketplace grows by pointing it at new corpora.
Every regulator: LTA → BCA, NEA, MOM, SCDF, PUB, PDPA. Same scaffold, new source docs.
Every internal corpus: QP policy library, Work Instructions, SDS, asset registers, tender archives.
Every existing card that can graduate: the current prompt cards (WI Reviewer, TMRB/TCCC/AFM checkers, Incident Drafter) are the natural Tier 1 candidates — they already have the logic, they just need the governed RAG layer.
After we launch the toolkit marketplace
The intent here is two-fold: find the blindspots we are missing today, and reduce adoption friction from senior management to the floor.
7a. Blindspots in the current design
1. The governance claim is unproven. The Depot says "governed, not shadow" — but a copy-paste prompt has no audit trail. If someone pastes a prompt with sensitive data into a public assistant, that is a data leak the current model cannot prevent or detect. This is the #1 blindspot. The marketplace must move from "here's a prompt" to "here's a governed agent with a real audit trail" for anything touching sensitive data.
2. No usage telemetry. The Depot has no way to know which agents are used, which produce wrong answers, which are stale. Governance without telemetry is blind. Need usage analytics + a "report a bad answer" button on every agent.
3. Version / currency drift. Regulatory codes change (LTA 2024 edition). A stale corpus gives confidently wrong compliance answers. Need "last updated" on every agent + auto-reindex when a source changes + a deprecation notice.
4. Hallucination in a compliance context is catastrophic. An agent that invents a clause is worse than no agent. Need citation-forcing (every answer must cite a clause), confidence gating, and human sign-off for sign-off decisions.
5. "One standard" is fragile. If agents are prompt-based, people can still improvise around them. The standard only holds if the governed agent is the path of least resistance — otherwise shadow IT grows back.
6. No access control. Compliance data is sensitive. Need role-based access: floor sees their WIs, managers see dashboards, compliance sees everything. The current model has none.
7b. Adoption friction, by level
Adoption is a change-management problem, not a technology problem. Each level has a different "what's in it for me" and a different objection.
| Level | Friction | Reduce |
|---|---|---|
| Senior management approve budget, carry liability | No business case, no risk sign-off, fear of liability. | A one-page ROI + a risk register + a named pilot sponsor + a "governance in a box" pack they can sign once. Show it reduces audit findings and sign-off time. |
| Middle management supervisors, engineers — gate daily use | "Is this replacing my team?", "do I trust it?", "who's accountable if it's wrong?", "extra process slows us down." | Show it makes their job easier (fewer audit findings, faster sign-off). Give them the accountability answer: the human signs off, the agent assists. Make it a tool, not a threat. Train them first as champions. |
| The floor technicians — decide whether it's used | "Another system to learn", "I don't trust AI", "it's faster to ask my supervisor", "no time." | Zero-training UX (paste + go), make it genuinely faster than the current path, show it catches things they'd miss (safety), start with the highest-pain task, celebrate wins, and give them a feedback loop so they feel heard. |
7c. The change-management layer
Named champions at each level — one senior sponsor, one middle champion, one floor champion.
A pilot with visible wins — pick one high-pain task, ship it, measure the before/after, publicise it.
Training that is role-specific — not "here's the tool," but "here's what it does for you."
A feedback loop — every agent has a "report a bad answer" path, and the fixes are visible. Trust is built by showing the tool listens.
Next 30 days
Five moves, in order. Everything else is cheap once the foundation exists.
1. Build the RAG scaffold (ingest → chunk → embed → retrieve → cite). This is the foundation; everything else is cheap once it exists.
2. Ship the LTA Code of Practice agent as the flagship — it is the reference implementation and the proof of the framework.
3. Stand up the governance-in-a-box pack — policy + risk register + approval template + test checklist, so Tier 1 agents fast-track.
4. Add telemetry + "report a bad answer" to the marketplace before scaling — governance without telemetry is blind.
5. Run the pilot — one high-pain task, one champion per level, measure before/after, publicise the win.