forgeloop 0.1.1

The glass-box agent loop — policy-governed AI agents with runtime gates, audit and escalation. GMS-optional.

Beyond Prompt and Pray

Building policy-governed AI agents from scratch.

Most agent code is "prompt and pray" — you wire up a model, hand it tools, and hope it stays in scope. This is the open, inspectable alternative: an agent loop you can read in an afternoon, with typed actions, runtime gates that detect and stop risky or out-of-scope behavior before it executes, multi-axis budgets, human escalation, and a tamper-evident audit log. Everything runs on the standard library plus pydantic — the baseline works with no license and no GMS.

For production-grade guarantees — geometric plausibility, calibrated admissibility, signed verdicts — the optional GMS backend (knowlytix) snaps in via a lazy seam. Clone it, run the baseline, see exactly where GMS lifts detection and validation.

Part of the "Beyond … and Pray" series: governed agents · trustworthy RAG · test & validate · LLMs from scratch

What's inside

• Governed agent loop — generator-based, every step inspectable
• Runtime gates — ALLOW / DENY / ESCALATE, stop bad actions before they run
• Typed actions — ToolCall / AskUser / Finish / Escalate
• Budgets — token / time / tool-call / dollar caps, first-class
• Human escalation — reviewer protocol, not an error path
• Tamper-evident audit — hash-chained record of every decision
• GMS-optional — baseline runs free; geometric guarantees via knowlytix

Install

pip install forgeloop                 # core: loop, typed actions, tools, gates, audit, planning, memory
pip install "forgeloop[ml]"           # + open-weight model tools (torch, transformers)
pip install "forgeloop[gms]"          # + the licensed GMS backend (knowlytix; see below)

Quickstart

A gate (or policy) is just (action, state) -> ALLOW / DENY / ESCALATE:

from forgeloop.core import ToolCall
from forgeloop.governance import pii_policy

action = ToolCall(tool_name="send_reply", arguments={"message": "Your SSN is 123-45-6789."})
verdict = pii_policy(action, None)
print(verdict.decision.value, "—", verdict.reason)   # escalate — detected PII: ssn

Those policies compose into a GovernanceHarness that runs gates before each agent action. For the full live ALLOW / DENY / ESCALATE stream and the hash-chained audit trail, see the runnable demos in beyond-ship-and-pray.

The GMS upgrade (open-core)

forgeloop runs fully without a license. The GMS-backed features — geometric plausibility gate, calibrated admissibility, signed verdicts — require the licensed knowlytix package, installed separately. forgeloop imports it lazily:

import forgeloop.gms as gms
gms.available()   # True if the licensed backend is installed
gms.require()     # returns knowlytix, or raises with install instructions

The production-grade, GMS-native edition is the Beyond Prompt and Pray, Pro Edition — see knowlytix.ai.

License

Apache-2.0. © 2026 Knowlytix.