FusionWork Academy has one job: help people hand real work to agents without giving up control.
The tools, models, and product screens will keep changing. The way you hand work to an agent should not. Every trustworthy handoff uses the same four parts:
- An outcome — what you want done, written in plain language.
- Steps in your existing tools — email, documents, chat, CRM, finance systems. The work does not move to a new app just because an agent is involved.
- Approval gates — a named person signs off before the agent sends, pays, deletes, or takes any other step you have marked as sensitive.
- An action log — what the agent saw, what it did, and who approved it.
If you remember only that pattern, you can judge a new model, a new connector, or a new vendor without starting from zero.
How much may the agent do alone?
Not every step deserves the same treatment. Before you hand a step to an agent, check four things:
- A clear boundary. You know the inputs, the outputs, and the person who owns the result.
- Data you can rely on. It is consistent, you can trace where it came from, and the agent is allowed to use it.
- A result you can check. You can measure it, test it, and correct it when it is wrong.
- Controls in place. There are limits, a record, a way to roll back, and a person to escalate to.
If any of the four is missing, fix that first. Then place the step on two questions: how certain is the task (stable rules, clear inputs, a right answer you can verify), and how bad is a mistake (who is affected, whether it can be undone, what the law expects).
| Mistakes are cheap and reversible | Mistakes are costly or hard to undo | |
|---|---|---|
| The task is certain | Run it automatically. Plain automation first. The agent handles messy inputs and the odd case the rules miss. Field checks, filing, a standard weekly report. | Rules lead. Fixed rules and system controls do the work. The agent spots exceptions and explains them; it does not replace the control. Payment runs, tax filings, stock adjustments. |
| The task needs judgement | Bounded autonomy. The agent drafts or acts inside limits, with sampling, a record, a cap, and an easy undo. Meeting notes, first-draft replies, internal questions. | People decide. The agent gathers evidence, compares options, and flags risks. A named person decides and owns it. Credit terms for a new customer, a large supplier contract. |
One rule overrides the grid: anything that leaves the company or cannot be taken back (sending, paying, deleting) counts as costly, so it always goes through a named person's approval. A workflow will usually mix all four boxes. That is normal.
Redesign, then hand over
Adding an agent to a step that is already broken just makes the broken step faster. Look at how the work really runs, then decide what to change before you hand it over: remove a step nobody needs, merge two handoffs, run independent checks side by side, move a check earlier, or keep a step with a person. The agent should take over a better workflow, not just the one written in the manual.
Five tracks, one pattern
| Track | Audience and completed work |
|---|---|
| A — Working with agents | Operations, delivery, sales, support and finance teams. Produce a four-part handoff with an access/privacy preflight and a named approver. No engineering background assumed. |
| B — Building trustworthy agents | Existing engineers converting to FDE work. Build and test a gated local workflow. Later volumes add identity lifecycle, integration, evaluation, operations, discovery and an assessed capstone. |
| Track | Audience and completed work |
|---|---|
| C — Build your own second brain | Founders and leads. Produce a source register, a compiled page with evidence, a scoped lint and a retrieval check under retention controls. |
| D — Running a multi-agent team | Founders and leads. Produce a small fleet map, separate access boundaries, an approval queue and an operating review. |
| E — Operator over OS (later) | Product and engineering leads and senior FDEs. Compare investment in workflow methods with investment in shared infrastructure, including counterexamples. |
All tracks use the four-part handoff. Track B provides implementation practice; the public local labs establish control behavior in a simulation. Real integration competence requires the separate supervised work described in the sign-off pathway.
Audiences we are writing for
The first readers are in Singapore and Thailand. Examples use teams that might keep a company in Singapore and suppliers, warehouses, or colleagues in Thailand, because that is a common shape of work in the region. When a module talks about personal data, it starts from Singapore's Personal Data Protection Act 2012 (PDPA). It also flags Thailand's Personal Data Protection Act (PDPA, B.E. 2562) when the work includes personal data of people in Thailand. This is a briefing for practitioners, not legal advice. Formal questions go to the customer's counsel and, for FusionWork's own site and services, to the Data Protection Officer.
Regional privacy and governance briefing
Citations are being verified against official texts; editions, dates and section references remain pending verification. Official starting points: Singapore Personal Data Protection Act 2012, Thailand PDPA — Ministry of Digital Economy and Society legislation index, and IMDA — Model AI Governance Framework for Agentic AI. These links do not mean citation verification is complete.
This is practitioner briefing, not legal advice. References: Singapore, Personal Data Protection Act 2012; Thailand, Personal Data Protection Act B.E. 2562; Singapore Infocomm Media Development Authority (IMDA), Model AI Governance Framework for Agentic AI. Citation review date: 16 November 2026, Asia/Singapore. For each reference, the official consolidated text/edition, publication or update date, URL and relevant section numbers are pending verification: citations are being verified against official texts; see the official source links in this briefing. Do not infer certification, statutory compliance or exact framework recommendations from this reference list.
Start with Singapore PDPA questions for the fictional Singapore company. Flag Thailand PDPA assessment whenever people in Thailand are involved; location alone does not settle applicability or the lawful processing/transfer route. Customer counsel and the customer's privacy lead decide those questions; FusionWork's own questions go to its DPO. A person approving an agent action is authorising that particular execution. It is not privacy consent from the people whose data appears in it. A connector permission grant is also not privacy consent.
Draw the data flow before access: source mailbox or file → local extraction → restricted evidence store → draft/approval view → destination connector → event record, snapshots, traces and backups. Mark the people, purpose, fields, readers, processing locations, remote access and service providers at every step. A Singapore mailbox viewed by a Bangkok operator, or Thai staff data processed in Singapore or another hosting region, raises transfer questions even when no attachment is emailed. Have the customer confirm the appropriate arrangements before using real data; no invented section number or generic consent checkbox resolves this.
Minimise records and set purpose-based retention, review dates, authorised access/correction/deletion procedures, and a backup-expiry schedule. Append-only logs and immutable source versions mean that ordinary agents cannot silently rewrite history; they do not mean keep personal data forever. Store minimal event metadata separately from restricted, expiring payloads. Controlled retention/correction can purge or redact source versions, snapshots, index entries, wiki pages, caches and backups, leaving only a minimal lawful audit marker. Hashes and tokenised identifiers may remain personal data if linkable. Holds need an authorised reason, owner, expiry/review date and restricted access; they are not permission to keep everything.
If something goes wrong, stop the affected workflow, restrict further exposure, preserve only appropriate evidence, and escalate to the customer's incident lead and privacy lead/counsel. Those roles decide containment, notifications and legal deadlines. The FDE does not independently notify people or authorities based on this briefing. The academy's rule that send, pay and delete always need separate named-person approval is academy policy, not a claimed statutory rule or a claimed IMDA prescription.
Minimum preflight for every initial release path
Before A1, B1, C1 or D1 invites action, record: authorised data and purpose; the smallest mailbox/folder/field/time access; a named owner and backup; where data flows and when it expires; the stop switch and person to escalate to; and the exact approval route. Start on fictional or explicitly authorised material. No live outbound action is allowed outside the stated approval process. If authority, translation, evidence or destination is uncertain, hold and ask the owner. A missing preflight means a paper exercise only.
A small glossary: three useful views
| View | Meaning and mapping |
|---|---|
| Four-part handoff | Outcome, steps in existing tools, gates, record. This is the owner's brief. |
| Four runtime components | Runtime/proposer, policy/gates, connector facade/adapters, action log. The outcome guides the runtime; steps use connectors; policy enforces gates; the log records intent and outcome. |
| Five-part harness | Memory (retained facts), context (current brief/evidence), verification (checks), tools (interfaces), guardrails (enforced limits). It is a practical operating setup across the four components, not five more services to install. |
| Session identity | A short description of an agent's role for one conversation. It grants no access and proves no identity. |
| Authenticated principal | An identity established by a trusted sign-in or mock authentication boundary and checked against current grants. Writing a person's name in a prompt cannot create it. |
| Grant, approval, consent | A grant permits access within scope; action approval authorises one complete execution envelope; privacy consent concerns use of a person's data and is a separate legal question. |