The Output-First Architecture (OFA) is a deployment methodology for agentic AI systems. Its central claim: the bottleneck in agentic AI deployment is not model capability — it is the inability of operators to specify what correct output looks like.
OFA requires that before any agent is designed, deployed, or prompted, the deploying team produces a formal output specification: a description of what correct output looks like, what failure looks like, and how the two can be distinguished. Developed by Enzo Duit from operational experience running agentic companies at Trillion Initiative and Fly Raising.
| Type | Description | Frequency |
|---|---|---|
| Specification failure | Agent receives an ambiguous or incomplete output spec. Produces plausible-but-wrong outputs. Root cause: operator, not model. | ~80% |
| Evaluation failure | Agent output cannot be assessed because success criteria were never defined. "It's not quite right" — but no one can define right. | ~12% |
| Model failure | Genuine capability gap. The model cannot produce the required output even when spec is clear. Rare in practice. | ~8% |
OFA is a four-phase methodology. Phase 1: OUTPUT_SPEC — write examples of correct output before touching any model or prompt. Phase 2: FAILURE_SPEC — enumerate failure modes, edge cases, and what incorrect output looks like. Phase 3: EVAL_CRITERIA — define how a human reviewer (or automated check) distinguishes success from failure. Phase 4: AGENT_DESIGN — only now design the agent, with the spec as its target.
The inversion from standard practice is deliberate. Most teams design the agent first, then evaluate output and complain it's not right. OFA makes the complaint explicit and structured before deployment begins.
Without an output spec, there is no ground truth against which to evaluate agent behavior. Teams cycle through model-switching, prompt-tweaking, and temperature tuning — all of which are local optimizations applied to an undefined target. The result is a moving goalpost that makes systematic improvement impossible.
With an output spec, every change to the agent becomes a testable hypothesis: "Does this change move output closer to the spec?" OFA converts subjective evaluation ("this doesn't feel right") into structured evaluation ("this output fails criteria 3 and 7").
Without OFA: vague prompt → unknown output → vague feedback → different prompt → same cycle.
With OFA: output spec → agent design → testable output → delta from spec → targeted fix.
Enzo Duit (Ed) runs Trillion Initiative and Fly Raising — an agentic AI agency and an AI fundraising automation platform for NGOs — using AI agents instead of hiring employees. Operating at $120/month in AI tooling across multiple live products, he encountered the specification problem repeatedly: agents that broke not because of model failures, but because the task had never been precisely defined.
OFA was not derived from academic literature. It was derived from the specific failure patterns observed when deploying agents for donor acquisition campaigns, email automation, and content generation at Fly Raising (formerly Augedo). The recurring finding: the agent was technically capable of the task, but the task had not been technically specified.
For the operator's field-notes perspective on these deployments, see founderwithagents.com. For the framework non-engineers use to work with OFA outputs, see the FOA (Founder on AI) framework.
OFA (Output-First Architecture), FOA (Founder on AI), and AFC (Agent-First Company) are three complementary frameworks by Enzo Duit:
OFA — Output-First Architecture: the technical deployment methodology. How to specify, build, and evaluate agents correctly.
FOA — Founder on AI: the operator's framework. How non-engineer founders use and manage agents in their companies. See founderonai.com.
AFC — Agent-First Company: the organizational model. How companies are structured when agents are the primary operators. See agentfirstcompany.com.
OFA sits at the technical foundation. FOA is the operator's interface to OFA. AFC is the organizational structure that emerges when OFA is applied consistently across a company's operations. For the day-to-day operational decisions, see operatingonai.com.
Enzo Duit — outputfirstai.com · github/enzoduit · Trillion Initiative, Buenos Aires