Loop Forge — Finta

01 · Agent identity

The bootstrap contract: who the agent is, what it may touch, and what it must never touch. Scope is locked at bootstrap — the agent cannot widen it mid-run.

Agent name

Lowercase, hyphenated, < 64 chars.

Target runtime

The architecture stays model-agnostic either way.

Mission — the verifiable objective

If success can’t be measured, the loop can’t know when to stop — the forge will flag it.

In scope (the agent may…)

Out of scope (the agent must never…)

02 · Loop profile

Three questions pick one of five loop profiles — and the profile changes how the runner executes: turn-by-turn runs one iteration per invocation, human-validated dry-runs until you --confirm, temporal waits on an event trigger, proactive keeps watching after success and re-enters on regression. Start simple; add autonomy only when the proof is reliable.

Is the task recurring?

Answer the questions above — or pick a profile directly.

03 · Guardrails

Emergency brakes, enforced in the runner — not documented in prose. The loop must always be guaranteed to stop.

Max iterations 20

Hard cap on automatic loop turns before HTL.

Stagnation — flat rounds before stop 3

N rounds without score progress → stop and escalate.

Doubt threshold 10%

Above this uncertainty, the agent must CLARIFY, not guess.

Max fix retries per bug 3

One diagnosis, one atomic fix per attempt; then BLOCKED.

Time budget (minutes, 0 = none)

Token budget (0 = none)

✓ Rollback on regression
If a verification that passed starts failing, restore the last healthy checkpoint before continuing.

04 · Human-to-loop triggers

Hard brakes — iteration cap, stagnation, budgets — always stop the loop; stopping is guaranteed and not configurable. These toggles govern the advisory triggers the agent raises itself ([HTL:…] markers): enabled → stop and escalate; disabled → logged warning, loop continues.

✓ Ambiguity
Context or objective is unclear — ask a targeted question instead of guessing. ✓ Irreversible action
Risky or impossible to undo — deploys, deletions, external sends. ✓ Sensitive data
Access to or processing of critical data requires a human OK. ✓ Subjective judgment
Quality calls a machine shouldn’t make alone — tone, design, tradeoffs. ✓ Budget exhausted — always on
Hard brake: time, iteration or token ceiling reached → stop and report. Not disableable — a loop must be guaranteed to stop.

05 · Verification criteria

“Goal verified?” is never one boolean. Deterministic criteria run a command and read the exit code. Soft criteria compare against a spec you validate once — after that the agent applies it autonomously.

Success threshold — weighted score to STOP 80%

Score ≥ threshold → RESOLVED. Below → new iteration. At the iteration cap → HTL.

06 · Orchestration

Strong coupling → one agent. Independence → fan-out. Dependence → sequential. Mixed → both. If decomposition benefit ≤ coordination cost, stay single.

Forge an agent that finishes the job.

01 · Agent identity

02 · Loop profile

03 · Guardrails

04 · Human-to-loop triggers

05 · Verification criteria

06 · Orchestration

Sub-agents

⟳ · Your loop architecture