Designing Agentic Systems That Survive Production

Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing cost, unclear value, and weak controls (Gartner, 2025). None of those reasons is the model. They are design problems, and this is how we design agentic systems to still be healthy a year after launch.

Why agents drift

An agent is a chain of probabilistic decisions that depends on four things, and all four change:

• Models. Providers update weights and deprecate versions.
• Prompts. Someone edits an instruction to fix one case and breaks three others.
• Tools. APIs gain fields, lose endpoints, tighten rate limits.
• Data. Next quarter's requests do not look like last quarter's.

The usual failure is not a crash. It is a slow fall in resolution rate that nobody sees, because nobody defined the metric.

Orchestration

Beyond a single agent with a couple of tools, we use explicit orchestration: a graph where each node is a model call, a tool call, or a deterministic check, and each edge is a named transition.

Two patterns have held up:

• A coordinator with specialist workers. The coordinator splits the goal and hands parts to narrow agents with narrow tool sets. Each specialist is easier to evaluate and cheaper to run than one generalist with forty tools.
• Deterministic gates around probabilistic steps. Prices, eligibility rules, and permission checks are code, not model judgment. The model proposes; the gate decides.

The pattern to avoid is one large prompt that does everything. It demos well and degrades without anyone noticing.

Evaluation comes first

Before building, we agree with the client what "resolved" means, in numbers. Those criteria become datasets of real scenarios with expected outcomes, and they run on every change: model upgrades, prompt edits, tool updates.

Without that harness you cannot change a model, or even reword an instruction, and know what broke. That is why evaluation criteria are written before code.

Tracing

Every production run gets a full trace: the goal, each model call with inputs and outputs, each tool call with arguments and results, latency, tokens, and cost. When something goes wrong, the answer comes from data, without reproducing anything.

Failures found in traces become new eval cases, so the same mistake is caught next time.

Cost control

A hard task can use fifty times the tokens of an easy one. Without controls the monthly bill is a surprise. By default we put in place:

• Per-task cost budgets with hard stops.
• Model routing: small models for classification and extraction, large ones only where reasoning pays.
• Caching of repeated retrieval.
• Cost per resolved task as a first-class metric, next to resolution rate.

After launch

Deployment is where the system's real life starts. Models, tools, and the business will keep changing. The systems that last are built for that: orchestration you can change piece by piece, evals that say what broke, traces that say why, and budgets that keep it affordable. It is also what lets a company move up the levels of autonomy on evidence instead of hope.

If you are designing a multi-agent system, or looking after one that drifted, that is our agentic systems work. Contact us and we will tell you where we expect it to break first.