Deterministic AI orchestration

Agents that run the same way every time.

Large language models predict the most likely next token. If you ask the same question 10 times you will get varying answers. Deterministic AI orchestration introduces guardrails that reduce the variance of large language model output for software development.

Get in touch Read what we learned building it

Scroll to learn more

The insight

Software development life cycle encoded — same input, same path, every run.

workflow.runDeterministic

● RUNNINGSetupWorkflow registration, triage, and git validation. The path is decided — and the run recorded — before any work begins.Exit gate

QUEUEDDiscoveryAll research runs through background agents on a research protocol. Read-only — structure and blast radius, before a line changes.Exit gate

QUEUEDDesignBlueprint, architecture documentation and implementation plan with a dependency graph and task queue, approved at a human gate before build.Human gate

QUEUEDBuildBuilt-in debate where every task gets a development and QA worker pair as background agents. Each completed pair ships an incremental PR and a review agent.Exit gate

QUEUEDVerifyCompliance checks, the full test suite, and coverage verification run before anything is allowed to ship.Exit gate

QUEUEDDeliverIncremental PRs merged, Layer 3 multi-select, and the outcome recorded back to the platform for the next run.Loop closes

Next task enters the same loop

The model

Thin agent. Fat platform. Thin skill.

Layer 01

Thin Agent

Spawned, scoped, and discarded. Holds no memory and makes no rules — it only executes the workflow it was handed. Disposable by design.

Layer 02 · Core

Fat Platform

The source of truth. State, knowledge, hooks, and codified workflows live here — versioned, enforced, and replayable. The platform is what makes runs deterministic.

Layer 03

Thin Skill

A routing skeleton, not a knowledge base. Skills point to the Tier-2 library where the real domain logic is codified once and reused everywhere.

An evolution of the platform architecture described in Praetorian’s “Deterministic AI Orchestration: A Platform Architecture for Autonomous Development”.

Principles

Five rules the platform enforces

Every rule is wired into hooks, scripts, and gates that run on every tool call.

Platform over improvisation

The platform enforces repeatable flows with an extensive set of protocols, hooks and scripts.

Hooks govern every tool call

Extended pre- and post-tool hooks are deterministic enforcement layers.

Agents are disposable; state is not

All persistent state lives on the platform. Kill an agent at any time — nothing is lost.

Context is a budget, not a workspace

Large corpora and multi-file work fan out to background agents, keeping the main context lean.

Roles are enforced, not assumed

Orchestrators coordinate and workers implement in developer-QA pairs.

Why ForgeShift

We built it, ran it, and learned where it breaks

Our own delivery platform runs on this architecture, so the lessons are not from a whitepaper: enforcement lives outside the model, context is a budget, and the agent is the disposable part.

01Enforcement, not documentation — rules that matter are hooks and gates that physically block the wrong action, not prose in a doc nobody enforces.

02Context as a budget — long research and multi-file work fan out to background agents, so the main run stays lean and repeatable.

03Disposable agents, durable platform — state, workflows and knowledge live on the platform, so any agent can be killed mid-run without losing the thread.

Read what we learned building it

Pitfalls we hit first, so you don’t

01 — Context rot

Long runs degrade quietly until the output stops being repeatable.

context-budget.yaml → compact at threshold
→ handoff.yaml + spawn background workers

02 — Markdown is a suggestion

A load-bearing rule written as doc prose is a rule the model may ignore.

"route every task" = a PreToolUse gate
that blocks — not a polite line in a doc

03 — Roles assumed, not enforced

Without enforced roles an orchestrator starts implementing, and the review step disappears.

PreToolUse: block foreground agents mid-run
orchestrator coordinates · workers implement

Every one of these was found by running the platform on real delivery work, not by designing it on a whiteboard.

Get in touch