Phase 9. Agents (Rust structs)

Field	Value
MEP	MEP-53 §Phases
Status	LANDED
Started	2026-05-28 (GMT+7)
Landed	2026-05-29 07:35 (GMT+7)
Tracking issue	— (umbrella)
Tracking PR	#22499
Commit	abf0aeeef0

Gate

TestPhase9Agents walks tests/transpiler3/rust/fixtures/phase09-agents/ (44 fixtures) and asserts byte-equal stdout. Coverage: agent declaration, agent spawn (spawn AgentType()), agent intent call, agent state mutation, multiple agents, agent with closures.

Lowering decisions

agent Counter {
    state count: int = 0
    on inc(by: int) { count = count + by }
    on get(): int { return count }
}

lowers to:

#[derive(Clone, Default, Debug)]
struct Counter {
    count: i64,
}

impl Counter {
    fn new() -> Self { Self::default() }
    fn inc(&mut self, by: i64) { self.count = self.count + by; }
    fn get(&self) -> i64 { self.count }
}

The key decision: agents are plain structs, not threads. spawn AgentType() lowers to AgentType::new() (immediate value construction); intent calls (a.intent(arg)) lower to direct method calls. There is no mailbox, no send / receive decoupling, no thread spawn. This matches MEP-53's single-thread runtime decision (§Abstract decision 2): users who need true concurrent agents can call into std::thread via FFI.

&mut self vs &self for intent methods is decided by the lower pass: if the intent body mutates any state field, &mut self; otherwise &self. The colour pass in phase 6 propagates the resulting borrow requirements through call sites.

Initial state values come from the state field: T = expr declarations and feed Default::default(). expr must be const-foldable; non-const initial values are rejected at lower time.

Files changed

File	Purpose
transpiler3/rust/lower/agent.go	Agent decl, spawn, intent lowering
transpiler3/rust/build/phase09_test.go	44-fixture gate
tests/transpiler3/rust/fixtures/phase09-agents/*.mochi + .out	44 fixtures

Test set

• TestPhase9Agents/<fixture> for each .mochi in the fixture directory (44 fixtures).

Closeout notes

The 44-fixture set is the largest in phases 0-14 because agents touch every prior phase: scalar state, list state, sum-type state, closures-as-state. Each fixture exercises one composition. Two fixtures were dropped pre-merge: one needed Rc<RefCell<Agent>> to share an agent between two intent callers (not yet supported), one needed Arc<Mutex<Agent>> for cross-thread sharing (explicitly out of scope per the single-thread runtime decision).