Accepting AI Work: The Five Gates Before Anything Ships

5 days ago
4 min read

An agent that can loop twenty times and correct itself is not, for all that, an agent authorized to ship alone. Between "it runs" and "it's in production," one thing is missing that the software industry has always known: acceptance.

We asked the question in a previous article: how do you prove those twenty turns were right? Here is the operational answer — the five acceptance gates our own pipelines (content, publishing, data) pass before anything goes live. None of them is theoretical; all of them run in our shop.

Key takeaways

A system that loops is not a system that's reliable; between the two sits an explicit acceptance process.
Five gates: deterministic checks first; the generator separated from the acceptor; evaluation sets per domain; actions classified reversible or irreversible; a human signature where legitimacy demands it.
The cost: 10 to 20% of pipeline time. The gain: a "done" that actually means something.

Gate 1: deterministic first

Anything a test, a compiler, a linter or a rule can settle is never handed to another model. It's the cheapest and most reliable gate: a link checker, a price-consistency check, a schema validator, a character counter — code, with an exit status that blocks the pipeline on failure.

The industry signal points the same way: SWE-bench, the coding-agent benchmark, went from under 10% to over 80% in a year — but what changed the trade is the word Verified: the result is confirmed correct by an external standard, not by the agent itself. Our rule states it plainly: never an LLM as judge of what a machine can settle.

Gate 2: the generator is not the acceptor

An agent's favorite failure isn't crashing, it's declaring victory too early: a plausible patch, a well-written report, "done" — and the critical path was never tested. A generator loves its own output; letting it grade itself is organized conflict of interest.

So our architecture separates the two roles: the agent that writes is never the one that validates. Acceptance runs in isolated subagents, with their own context, that return only a verdict — never their intermediate states. The main thread stays clean, and the verdict isn't contaminated by the writer's enthusiasm.

Gate 3: evaluation sets per domain

How do you know a prompt or model change broke nothing? The same way software does: regression tests. An evaluation set is 30 to 100 cases with known-correct outcomes, specific to each domain — content compliance, data extraction, code fixes. Run it before every change; drift shows up as numbers, not impressions.

For content, we stage acceptance in four levels: facts (are the numbered claims right?), links (do they resolve?), compliance (the domain's required mentions and prohibitions), brand voice (no emoji, prose before bullets, numbers before adjectives). A text ships only when all four levels pass.

Gate 4: classify actions — reversible or irreversible

Not all agent actions are equal. Writing a draft is reversible; publishing under the brand's name is not. Writing to staging is reversible; touching the production database is not. Our rule: the agent moves freely on the reversible; the irreversible passes a gate — restricted tool permission, budget circuit-breaker, or explicit approval.

This is runtime governance in miniature: the loop gives the agent the power to act, the classification gives it boundaries. Most agent accidents we hear about trace back to the absence of this gate — not to a weak model.

Gate 5: the human signature, where legitimacy demands it

Some domains keep a human in the loop whatever the model's quality: healthcare, finance, legal — because law and accountability require someone who signs. And beyond regulation, anything that goes out publicly under the company's name deserves a final human read, at least at first; sampling gradually replaces systematic review as the evaluation sets accumulate trust — never on anything compliance-critical.

This gate is not an admission of technical weakness. It's a matter of legitimacy: an agent can be accurate and still not be authorized.

What it changes

The five gates cost 10 to 20% of pipeline time. In exchange, "done" rests on tests, independent verdicts, regression numbers and action boundaries — not on the self-report of a system that loves to conclude. Our conviction, stated elsewhere, proves itself concretely here: the loop is a commodity, the verifier is the moat. Models will keep making the loop cheaper; acceptance remains your job.

FAQ

Do you have to accept everything at 100%? No. Classify by risk: the irreversible and compliance-critical get 100% acceptance; low-stakes reversible work gets sampled. Gate 4 exists precisely to establish that triage.

Can an LLM serve as judge? Yes, for what no test can settle — tone, clarity — provided it's a separate agent with an explicit rubric. Never for what a machine can verify.

Where to start? Inventory your agents' outputs, classify them reversible/irreversible, then install Gate 1 this week: deterministic checks are the cheapest and pay off immediately.

ECTIME AI Lab is the applied-AI research and deployment unit of ECTIME Group. We build, ship and stress-test agentic systems in production, from GEO/SEO automation to multi-step autonomous agents. We maintain open-source Claude Skills for GEO/SEO and advise European brands on deploying AI that is not just autonomous, but verifiable and authorized.