Architecture Decision Map for Agentic Systems

Agent architecture is one decision, not a series of separate ones. It settles how control flows, what the model is for, which tools exist, where state and memory live, who holds authority, how a person steps in, and how the result is judged. Choose only a framework or a model and the most important choices stay unmade.

The right architecture can be a direct model call, a deterministic workflow, a bounded agent, a long-running durable process, or a multi-agent system. Complexity must be justified by the task.

Authority is the column a team least often gets to fill in freely. For high-risk systems the law fills part of it in before the design starts. The EU AI Act entered into force on 1 August 2024. Its Article 14 requires that high-risk systems be designed so that natural persons can effectively oversee them in use. The assigned overseers must be able to decide not to use the system, and to disregard, override or reverse its output. They must also be able “to intervene in the operation of the high-risk AI system or interrupt the system through a 'stop' button or a similar procedure that allows the system to come to a halt in a safe state.”

Read that as an architectural constraint, not a compliance note. A halt in a safe state has to be reachable from every point the loop can occupy. That bounds how much uncommitted state the system may accumulate, how far a tool call may go before it turns irreversible, and how many agents may hold pieces of one task at once. Pick the loop, the tools and the memory first, ask about the stop button afterwards, and all three reopen.

Start with what a free-running agent actually completes on realistic multi-step work. WebArena is a reproducible environment of fully functional websites in four domains, and frontier models were turned loose in it. The 2023 result: “our best GPT-4-based agent only achieves an end-to-end task success rate of 14.41%, significantly lower than the human performance of 78.24%.” On long-horizon work in a realistic environment, the unconstrained agent finished about one task in seven. People finished about four in five. That is the baseline any broad-access planner has to be argued against. It makes a bounded design that wins on accuracy, cost and auditability the expected outcome rather than a lucky anecdote.

The harder question is whether adding agents buys anything. June 2025 answered it twice, in opposite directions, one day apart. On 13 June Anthropic reported that a multi-agent system beat a single agent by 90.2% on its internal research evaluation. The lead agent was Claude Opus 4, the subagents Claude Sonnet 4; the beaten baseline was Claude Opus 4 alone. The same post priced the arrangement: “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats”. The day before, on 12 June, Cognition's Walden Yan had argued in Don't Build Multi-Agents that such architectures are very fragile, because decision-making becomes too dispersed and context is not shared thoroughly enough between agents.

Neither post is wrong. They measure different tasks, with different tolerance for dispersed decisions, and both are stating a price. A 90.2% gain bought at roughly fifteen times the tokens of a chat is a trade, and a trade can be a bad one. A fragility argument is a claim about which tasks cannot absorb that dispersion. Neither of them licenses treating coordination as an upgrade a serious system is simply expected to have.

Task uncertainty, action consequence, horizon, and evidence quality are the four readings an architecture choice rests on. Evidence quality and coordination need should not share an owner or a test. Each reading has a published measurement behind it, so none of them has to be guessed.

Task uncertainty belongs to the task distribution, not to how hard the task feels to the designer. The GAIA benchmark posed 466 real-world assistant questions in 2023, requiring reasoning, multi-modality, web browsing and tool use. The gap turned up exactly where intuition says there should be none: “GAIA questions are conceptually simple for humans yet challenging for most advanced AIs: we show that human respondents obtain 92% vs. 15% for GPT-4 equipped with plugins.” A task a designer can describe in one sentence is not thereby a task with low uncertainty.

Action consequence asks how reversible, costly, or sensitive the effects are. For high-risk deployments the EU AI Act has already constrained it, through the override and the safe-state stop of Article 14.

Horizon has a unit, and the unit is expert-minutes. METR timed domain experts on RE-Bench, HCAST and 66 new shorter tasks to define a 50%-task-completion time horizon, and reported in 2025: “On these tasks, current frontier AI models such as Claude 3.7 Sonnet have a 50% time horizon of around 50 minutes. Furthermore, frontier AI time horizon has been doubling approximately every seven months since 2019, though the trend may have accelerated in 2024.” Measure your own task in expert-minutes. Then compare it against the length a model finishes at coin-flip reliability.

Evidence quality is where one passing trace does the most damage. The pass^k metric, introduced by τ-bench in 2024, credits an agent only when it succeeds on all k independent trials of the same task. What it found: “Our experiments show that even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail).” Agents that pass roughly half of tasks once pass fewer than a quarter of them eight times running. A demo is evidence of nothing except that the run existed.

Coordination need asks whether one agent, several specialists, or humans must share the work. Five orchestration patterns for agents are documented in Microsoft's Azure Architecture Center: sequential, concurrent, group chat, handoff and magentic. That is a menu, not a ranking. Picking from it is the decision the other four readings are supposed to inform.

Write the architecture decision record for a system your team has already chosen. Write it as though defending that choice to someone who wanted a smaller one. It should expose “Framework features can disguise an unjustified architecture” and preserve evidence for “Each architecture component addresses a named failure mode”. One paragraph per component, each naming the failure it exists to catch.

The naming does not have to come from imagination. A 2025 paper asking why multi-agent LLM systems fail assembled MAST-Data: 1,600+ annotated execution traces across 7 popular multi-agent frameworks. Out of it came MAST, the first Multi-Agent System Failure Taxonomy. The paper describes how: “We develop MAST through rigorous analysis of 150 traces, guided closely by expert human annotators and validated by high inter-annotator agreement (kappa = 0.88). This process identifies 14 unique modes, clustered into 3 categories: (i) system design issues, (ii) inter-agent misalignment, and (iii) task verification.” Fourteen named modes in three categories is a checklist a reviewer can hold up against a design.

The record then follows five moves. Describe the task distribution, including easy, ambiguous, adversarial, and exceptional cases; GAIA is the warning that the easy-looking cases are not the safe ones. Score the four drivers: uncertainty, consequence, horizon in expert-minutes, and evidence quality. Choose a minimal baseline, and implement the least agentic option that can plausibly succeed. Add bounded mechanisms only for measured gaps, one at a time, each paired with the MAST mode or other named failure it exists to catch. Last, set a decommission rule for when cost, incidents, or weak evidence require simplification. State it in the units the evidence comes in — a pass^k floor, not a passing demo.

A component whose paragraph cannot name its failure mode is the component to remove first.

The architecture decision map closes the foundation module. Every later mechanism should be evaluated against the task contract and the simplest credible baseline.

The reviewer who signs off on agent architecture selection owes the room two things: the failure “Framework features can disguise an unjustified architecture”, and the evidence “Each architecture component addresses a named failure mode”. Both are now checkable against published numbers rather than against taste. Does the system run longer than the tasks a frontier model finishes at 50% reliability — around 50 minutes of expert work, on METR's measurement? Then the horizon column needs an answer. Is the evidence for it a trace that passed once? τ-bench's pass^8 under 25% in retail says what that trace is worth. Did the design add agents? Anthropic's 15× token figure and Cognition's fragility argument are the two halves of the price it paid, and the record should say which half the team accepted. Is the deployment high-risk under the EU AI Act? Then the override and the safe-state stop of Article 14 are not negotiable additions to that design.

And if the answer to any of these is that the framework came with it, the review has not happened yet.