Why most agent projects stall
The demo is never the hard part. A convincing agent demo takes a weekend. What kills agent projects in production is the three things a demo never has to answer:
- What is it allowed to do alone? An agent that can send an email, quote a price, or commit a date is making binding representations on the company's behalf. Most projects never define that boundary and end up either paralysed or dangerous.
- What does it remember? Agents that "learn" by appending everything to their prompt get more expensive and less accurate every week. Memory is several different things and they belong in different places.
- Who reviews the output, and can they keep up? Agent systems fail from approval saturation far more often than from bad reasoning — drafts arrive faster than any human can adjudicate, everything gets rubber-stamped, and the safety model becomes fiction.
We design for those three questions before writing agent logic, because they determine the architecture.
How we build agent systems
Draft, don't send
Anything crossing the company boundary — a message, a price, a date, a spend — is drafted and held for human approval. Not a training-wheel phase; a permanent design property.
One unit of work
An agent's only output is a single approvable action with its evidence attached. If a capability can't produce one, it isn't built yet. This is what stops agent fleets becoming prose generators.
Event-triggered, not polling
Agents wake when state changes, not on a timer. Timer-driven fleets burn budget continuously and generate alerts whether or not anything happened.
Cost as a column
Token spend books against the work item that caused it, so cost per outcome is a number you can query — not a surprise on the monthly invoice.
Bounded memory
World state in the database, episodic history append-only, learned rules capped and human-editable, prompts version-controlled in git where they can be diffed and rolled back.
Queue back-pressure
A hard ceiling on open drafts. An agent that can't add work must first close something — which forces prioritisation into the system instead of onto a person.
The proof is that we run on it
Valkyn's own operations run on this architecture: agents draft content and triage inbound, and a person approves anything that reaches a customer. We are extending the same system to proposal preparation and delivery admin.
That is not a marketing line — it is why we are opinionated about the boring parts. The alert-flood problem, the unbounded-prompt cost problem and the approval-saturation problem are all things we hit in our own system before we had a view about them.
If you want to see it, ask. It is more useful than a demo built for the meeting.
Where agents genuinely pay
Agents are worth it where the work is high-volume, judgment-light, and currently done by someone expensive between other tasks:
- Document-heavy intake — tenders, contracts, specifications: extract, classify, route.
- Drafting from an approved corpus — responses, proposals and replies grounded in what the company has already said.
- Triage and routing — deciding who should handle a thing, and escalating when the answer is unclear.
- Follow-up discipline — the status updates, reminders and chase-ups that get dropped when everyone is busy.
They are a poor fit where the work is low-volume and high-consequence, or where a deterministic rule would do the job more cheaply and more reliably. We will say so.
Common questions
What is an AI agent in a business context?
An AI agent is software that takes a goal, breaks it into steps, uses tools or data sources to act, and produces a result — rather than answering a single prompt. In a business setting the useful definition is narrower: an agent is a component that proposes a specific action with supporting evidence, which a person or a rule then approves or rejects.
Should AI agents be allowed to act without human approval?
It depends entirely on whether the action crosses the company boundary. Research, internal scoring, drafting and internal checks are safe to run unsupervised. Anything that sends a message, commits a price, promises a date, or spends money creates a binding obligation and should require human approval — regardless of how accurate the model is, because accuracy is not the same as authority.
How much do AI agent systems cost to run?
Ongoing model cost is usually far smaller than teams expect, provided agents are event-triggered rather than running on timers, and provided accumulated context is capped. The common failure is an agent whose prompt grows continuously as it 'learns' — cost rises every week while quality falls, because the added context is mostly noise.
How long does it take to build a production AI agent system?
A single well-scoped agent with a human approval step typically reaches production in three to six weeks. Multi-agent systems take longer, but usually not because of the agents — the time goes into defining what each is permitted to do alone and building the approval path that a real person can sustain.
What is the biggest reason AI agent projects fail?
Approval saturation. The agents work, but they generate proposed actions faster than anyone can review them, so reviewers begin approving in bulk and the human oversight becomes nominal. The structural fix is back-pressure — a hard limit on how much unreviewed work can exist at once — not better models.
Can agents work with our internal data without it leaving our systems?
Yes. Agent systems can be deployed inside a client's own infrastructure, with model access configured so that submitted content is not retained or used for training. For regulated or commercially sensitive workloads this is usually a requirement rather than a preference, and it should be settled before architecture, not after.
Describe the process that costs you the most
If a custom build is not the right answer, that is a short conversation and it costs you nothing.
info@valkyn.ai