An agent working through a task often reaches a question your organisation already has rules for, such as whether a refund is due. Discover lets it find the Decision your team built for that judgement and call it, instead of working out an answer of its own.
Only Decisions a person has offered can be found, and every search is recorded with the calls it led to.
- Who
- Agents in Claude Code, Codex or any MCP client
- Where
- The TypeFuse MCP server, with three tools
- You end up with
- Agents using your Decisions for the judgements they cover
Offer it when it’s ready.
A Decision stays private to your team until a person offers it. Its summary tells agents what it’s for and when to use it, in words a person wrote and checked.
- Off by default, for every new Decision, unless your organisation chooses otherwise
- A purpose and “when to use” are required before it can be offered; the assistant can suggest both
- Plain words only. Text that reads like an instruction to an agent is refused
acme/refundsSummary- Purpose
- Checks whether a customer refund request is eligible, and where it should go.
- When to use
- A customer asks for their money back on an order.
- When not to use
- Exchanges, or refunds a person has already approved.
- Example intents
- refund eligibilitycan I get my money backdamaged item refundlate refund request
Search for the judgement.
Agents search for the kind of judgement they need, and get the best matches with the exact call to make. The details of the case never go into a search: they go into the call.
- Nothing it can’t call. Results never include a Decision the agent isn’t allowed to use
- Searches with no match are recorded, so you can see which Decisions are missingSoon: not available yet
refund eligibility for a damaged itemJane Doe, order 88412, £84, delivered 12 days ago, says it arrived brokenacme/refunds1.4.0A customer asks for a refund and you need to know if it’s eligible.
call_decision ·acme/refundsacme/returns1.2.0Choosing the return steps for an item.
acme/disputes2.0.1A card payment is disputed with the bank.
Read the contract.
Before its first call, an agent can read exactly what the Decision needs and what it answers, from the live release.
- The inputs it needs, straight from the Definition’s
state_schema - The answers it gives, in TypeSafe’s native shape
- Which release is live, so the agent knows what it’s calling
{
"decision": "acme/refunds",
"release": "1.4.0",
"purpose": "Checks whether a refund request is eligible…",
"input": {
"days_since_delivery": "integer, required",
"item_condition": "new | used | damaged, required"
},
"answers": { "route": "choice: approve | review | decline" }
}Find once, call many.
An agent searches once for each kind of judgement, then calls the Decision for every item. When a job becomes routine, it moves into your code with the Decision bound by name.
- One search, many calls, each with its own sealed record
- Remembered across sessions when you add the name to your agent’s instructions
- Moves into code when the job is routine: your app calls the Decision by name, with no search
- search_decisions“refund eligibility…”3 matches
- describe_decision
acme/refundsinput and answers - call_decision#1042 · 12 days · damagedreview
- call_decision#1043 · 3 days · newapprove
- call_decision#1044 · 40 days · useddecline
One search, then a call for each refund. Put acme/refunds in your AGENTS.md and the next session skips the search.
See what agents looked for.
Every search is recorded with the calls it led to, so you can see which Decisions agents rely on and which judgements they couldn’t find.
- Searches, matches and calls, linked from one to the next
- Searches with no match, grouped, as a list of Decisions you might add
- Private by default: the words of a search are kept only if you choose
refund eligibilityacme/refunds212 callsreturn steps for an itemacme/returns64 callswarranty claim for a broken itemNo match14 searchesA search that finds nothing may be a Decision you don’t have yet. Search text is kept only if you turn it on; otherwise TypeFuse keeps the shape of each search, never its words.
Where it happens.
In the app
- Write the purpose and when to use it, or let the assistant suggest them
- Turn “Offered to agents” on for each Decision
- See searches and the calls they led toSoon: not available yet
Your agent, over MCP
search_decisionsfinds Decisions for a judgementdescribe_decisionreturns what one asks for and answerscall_decisionasks it, and the answer is recorded
Your code
- A routine job binds the Decision by name with the SDK
- No search at run time:
runtime.decide("acme/refunds", …) - Same record, same release rules
| Tool | What it does | Changes anything? |
|---|---|---|
search_decisions | Finds offered, deployed Decisions for a judgement, each with the exact call to make | No |
describe_decision | Returns one Decision’s purpose, input, answers and live release | No |
call_decision | Asks the Decision; returns its answers, the release and the record’s id | Records a call |
A Decision that isn’t offered, or isn’t deployed where the agent may call, is refused by name and never swapped for another. The call checks again even if the search was a moment ago.
Private preview
Own your next decision.
Keep the judgement calls your agents and apps make under your control, with a record of every answer.
- Your criteriaThe people who own a decision set its rules, and change them when the business changes.The people who own a decision set its rules.
- Your evidenceEvery release is tested against real cases before your application ever sees it.Every release is tested on real cases first.
- Your recordEvery answer is traced and sealed, ready for an auditor or the customer who asks why.Every answer is traced and sealed.