← Back to Autonomy
Field Notes · Agentic AI · June 2026

Scaling AI agents sounds easy — until everything breaks

Why a single impressive agent demo is the easy part, why errors in agent systems snowball instead of staying put, and what a production-grade architecture actually needs. In simple words, with practical notes.

By Majid Mazouchi Based on: IBM Technology — Sam Anthony[1] Evidence: arXiv 2512.08296[4] Reading time: ~8 min
Volume II — The Production Playbook → The build manual: context engineering, tool design, orchestration patterns, reliability, security, evals, and the pre-launch gate.
The one-paragraph version

TL;DR

Building one agent that does one task in a demo is easy. The hard part begins when you give agents more scope, more tools, more autonomy, and more users. Unlike normal software, an agent's mistake doesn't stay where it happened — it flows into every step that follows. The fix is not a better prompt. It is architecture: small bounded agents, verification gates, traceable memory, and humans at the irreversible moments.

Practical note

If you remember one sentence: agentic AI becomes dangerous or useless when autonomy grows faster than verification. Every design decision below is a way of keeping those two in balance.


First, a definition problem

"Scaling" means two different things

When people say "we need to scale our AI agents," they usually mix up two very different problems. Only one of them is solved by adding servers.

Meaning 1 — Scaling traffic

Same agent, more requests

More users, more calls per second, bigger queues. This is classic infrastructure work: load balancing, caching, rate limits, cost control.

Hard, but a known problem. Cloud engineering has solved this for 20 years.

Meaning 2 — Scaling capability

Same agent, more responsibility

More tools, more domains, deeper plans, longer memory, bigger decisions. The agent's scope grows — and with it, the number of ways it can go wrong.

This is the new problem, and the subject of this page. It changes the failure modes, not just the load.
Practical note

In planning meetings, ask which meaning is on the table. "We need agents to handle 10× more tickets" and "we want the agent to also issue refunds" sound similar but need completely different engineering. The second one needs a safety review, not a bigger GPU.


How agents actually work

The agent loop — and where it cracks

Under the hood, almost every agent runs the same simple cycle. It reads the goal, makes a plan, takes an action (often by calling a tool or API), stores what happened in memory, reflects on the result, and repeats until it decides it's done.

PLAN ACT call tools / APIs MEMORY store context REFLECT repeat until "done" — the agent itself decides when that is ⚠ a wrong assumption here… …is still trusted here
The basic agent loop. It works beautifully for narrow, bounded tasks. The crack: nothing in the loop questions the plan's original assumptions. Whatever the agent believed at step one gets carried — and amplified — around every lap.

For a small, well-bounded task ("summarize this document," "check this signal against this threshold") the loop is reliable. The trouble starts when the same loop is given a broad goal, many tools, and freedom to run for many steps.


The core failure mode

Errors don't stay put — they cascade

In traditional software, a bug is usually local: one function returns the wrong value, one test fails, you fix it. In an agent system, a wrong assumption made early becomes the foundation for everything after it. The classic illustration is a travel-booking agent told to plan a trip to "Washington":

SEEDAgent assumes "Washington" = Washington, D.C. (user meant Washington State)
STEP 2Books flights to the wrong coast — and they look perfectly valid
STEP 3Reserves hotels near the wrong landmarks
STEP 4Builds a day-by-day itinerary around the wrong city
STEP 5Computes budget and schedule — all internally consistent, all wrong
MEMORYStores "user is traveling to D.C." — future sessions inherit the error

Notice what makes this nasty: every individual step succeeded. No API failed, no exception was thrown, every output looks professional. The system is confidently, coherently wrong — and the error even gets written into memory, where it can poison future conversations.

Practical note

Three defenses, in order of value: (1) confirm ambiguous inputs before acting ("D.C. or the state?" costs one turn and saves the whole run); (2) checkpoint before irreversible actions — booking, paying, deleting, writing to a database; (3) store memory with provenance — record why the agent believes something, so a bad belief can be traced and removed instead of silently trusted forever.

Engineering caution

"Every step passed" is not the same as "the run is correct." Per-step success metrics will tell you this system is healthy. You need end-to-end, outcome-level evaluation — did the user actually get a trip to the right Washington? — or your dashboards will be green while your users are furious.


Why "one big agent" fails

The super-agent bottleneck

The tempting design is one powerful agent that does everything: understands intent, plans, picks tools, calls APIs, interprets results, manages memory, recovers from errors, and decides when it's finished. For a demo, this works. At scale, it's like running a company where one person makes every decision — every choice queues behind a single mind, and one bad judgment contaminates everything downstream.

Each added tool or domain multiplies the combinations the agent must handle. Context windows fill with irrelevant history. Latency and cost grow with every reasoning step. And when something goes wrong, you can't tell which of the agent's twelve responsibilities failed.

The remedy is the same one software engineering already knows: separation of concerns. Split the monolith into small agents with clear jobs and clear boundaries:

PLANNER decomposes the goal RETRIEVER facts, deterministic EXECUTOR bounded tool calls SPECIALIST domain reasoning MEMORY with provenance VERIFIER GATE checks outputs before they propagate HUMAN irreversible steps only each box is small, testable, and replaceable — failure stays local
Modular agent architecture. The teal verifier gate is the heart of the design: nothing flows downstream — and nothing irreversible happens — without passing a check. Research on agent scaling finds that architectures without centralized verification propagate errors significantly more than those with it.[4]
Practical note

A useful division of labor: deterministic services for facts, LLM agents for reasoning, verifier gates between them. If a value can be computed or looked up exactly (a parsed file, a database row, a graph traversal), never ask a language model to "remember" it — pipe it in as ground truth and let the agent reason on top of facts, not about them.


Not just opinion

What the measurements say

A 2026 study tested 260 agent-system configurations across six benchmarks and five architectures (single-agent plus four multi-agent patterns) to measure when adding agents helps and when it hurts.[4] Three results are worth keeping on a sticky note:

+80.8%

Multi-agent gain over single-agent on decomposable tasks (e.g. financial reasoning that splits into independent parts).

−70.0%

Multi-agent loss on sequential planning tasks, where coordination overhead outweighs the benefit.

Verify > Vibe

Architectures without centralized verification propagated errors more than those with it, across the board.

The headline: more agents is not automatically better. Architecture must match the task. Tasks that decompose cleanly love multi-agent designs; tightly sequential tasks often run best with one careful agent plus a verifier. Coordination also shows diminishing returns once a single agent is already strong at the task.

Engineering caution

Don't adopt a multi-agent framework because it's fashionable. Ask: does my task decompose? If every step depends on the previous one, splitting it across agents adds handoff errors and cost while removing context. Measure before and after — agent quality is harder to validate than normal software output, which is exactly why evaluation loops must exist from day one.


The takeaway, as a checklist

Six rules for production agent systems

  1. Bounded agents, not one giant agent Each agent gets a narrow job, a short tool list, and explicit permissions. Scope is a security boundary, not a convenience.
  2. Explicit ownership of planning, retrieval, execution, and verification If you can't name which component is responsible for a failure, the architecture is wrong. Separation of concerns is what makes debugging possible.
  3. Checkpoints before irreversible actions Booking, paying, writing, deleting, sending — anything that can't be undone gets a gate: a verifier agent, a confidence threshold, or a human. Assist agents recommend; humans decide on high-stakes calls.
  4. Memory with provenance, not loose conversational memory Every stored belief carries its source and timestamp. When a belief turns out wrong, you can find it, trace what it touched, and evict it — instead of wondering why the agent keeps repeating an old mistake.
  5. Evaluation loops from day one Measure outcomes end-to-end, not just per-step success. Build regression suites of past failures. An agent system without evals is a demo, regardless of how it's deployed.
  6. Fallback and handoff when confidence drops The agent must be allowed to say "I'm not sure" and escalate — to a simpler deterministic path, to another agent, or to a person. A graceful handoff is a feature; a confident wrong answer is an outage.
Practical note — applying this to engineering R&D

For diagnostics, prognostics, and knowledge-engine work, the pattern translates directly: keep data extraction and parsing deterministic (signal pipelines, file parsers, graph traversal are code, not prompts); keep facts in a structured store the agent reads but cannot rewrite freely; use LLM agents only where reasoning, synthesis, or explanation is genuinely required; and put a verifier between agent output and anything a downstream engineer or system will act on. The scaling problem is mostly system architecture, not prompt engineering.

Continue to Volume II — The Production Playbook → Nine engineering disciplines that implement these six rules: context engineering, tool contracts, orchestration patterns, hard stops, the lethal trifecta, evals, versioning, cost control, and observability.

Sources

References & further reading

  1. Sam Anthony (IBM Technology), “Building AI Agent Systems and Scaling Challenges in Agentic AI,” YouTube, June 2026. The talk this page is based on. youtube.com/watch?v=fCHe_fOqlYA
  2. IBM Think, “Building and evaluating AI agents that work in the real world,” May 2026. Source for the assist-vs-decide principle, agentic AI as an architectural layer, and scaling pressures (coordination, consistency, traceability). ibm.com/think/insights/building-evaluating-ai-agents-real-world
  3. IBM, “AI agents: Opportunities, risks and mitigations” (white paper) and the companion article “Scaling Responsible Agentic AI,” Nov 2025. On new and amplified risks of agentic systems, including irreversible actions and untraceable behavior. ibm.com/think/insights/scale-responsible-agentic-ai
  4. Kim, Y. et al., “Towards a Science of Scaling Agent Systems,” arXiv:2512.08296, 2026. 260-configuration empirical study: +80.8% / −70.0% architecture-task results, capability saturation, and the error-propagation advantage of centralized verification. arxiv.org/abs/2512.08296
Honesty note

The video transcript was not directly accessible at the time of writing. The framing, the two meanings of scaling, the agent loop, the cascading "Washington" example, and the single-agent bottleneck analogy follow the talk's published description and consistent IBM teaching material; treat the worked example as illustrative of the talk's argument rather than a verbatim quote.