Architecture Fieldnotes

Interactive reference · Version 2026-09-18.1

AI Agent Architecture
Pattern Selector

Start with the workload. Choose a candidate architecture, inspect its disqualifiers and export the reasoning for your design review.

01 / Your constraints

Example: read-only research. Adjust it to your workload.

Decision inputs
Workload structure
Runs locally. No model or form-data upload.

02 / Architecture candidate

EXAMPLE · LOADING RULES

ReAct loop

The example discovers actions from observations. Enable JavaScript to change the inputs and generate a decision.

Why this result

  • The interactive result will include every conflict and its rationale.

Required review checks

  • Validate outputs, scope access and record material execution events.
Compare alternatives ↓

Architecture guidance only. Validate the design against your workload; this result is not deployment approval or a guarantee of model behavior.

Every alternative, with a reason

Compatible alternatives can lose on rule order.

Selection and exclusions for the current inputs. No numeric fit score.
PatternDecisionReason
Enable JavaScript to compare your constraints. Static pattern references follow.

The pattern reference

Six primary choices and one optional feedback overlay.

Baseline

Single model call

One bounded model task with an input contract and a checked output.

  1. Input contract
  2. One model call
  3. Validate output
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
Extraction, classification or drafting that needs no model-directed tool loop.
Reconsider when
The next action depends on a tool observation, or the task needs several model steps.
Bring to review
An input/output schema and a small representative acceptance set.
S01 sourceS02 source

Defined sequence

Fixed workflow

Application code owns several model steps and their allowed transitions.

  1. Validate input
  2. Configured steps or branches
  3. Merge and validate
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
A known sequence, router or configured fan-out with explicit merge rules.
Reconsider when
The task requires the model to invent new steps or delegates during execution.
Bring to review
A transition table with branch conditions, merge rules and failure exits.
S01 sourceS02 source

Observation-driven

ReAct loop

Reasoning, actions and observations alternate as the task unfolds.

  1. Choose action
  2. Call scoped tool
  3. Observe and stop or repeat
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
A tool-using task where the next useful action depends on the latest observation.
Reconsider when
You cannot bound the available tools, iterations or stopping condition.
Bring to review
An allowed-tool contract and a trace showing stop, failure and budget exits.
S03 source

Planned decomposition

Plan and execute

A planner creates steps, an executor performs them, and a later check may trigger replanning.

  1. Plan steps
  2. Execute a step
  3. Check completion or replan
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
A decomposable goal whose steps can be reviewed before execution.
Reconsider when
The initial plan cannot name useful steps, or replanning may bypass required approval.
Bring to review
A versioned plan with dependencies and a reapproval rule for plan changes.
S06 source

LLMCompiler-style

Dependency-aware execution

A task graph makes dependencies explicit so independent function calls can be scheduled together.

  1. Plan dependency graph
  2. Run ready independent tasks
  3. Join results
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
A planned task with demonstrably independent branches and a defined join.
Reconsider when
Hidden shared state or output dependencies make the proposed parallel work unsafe.
Bring to review
A dependency graph with read/write sets, join semantics and partial-failure handling.
S04 sourceS02 source

Dynamic delegation

Supervisor and specialists

A coordinator delegates work to separate specialist contexts and integrates their results.

  1. Coordinator assigns
  2. Specialists work in scoped contexts
  3. Coordinator integrates
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
Dynamic work requires distinct tools, context boundaries or specialist responsibilities.
Reconsider when
Delegation adds no necessary separation, or coordinator authority and handoffs remain undefined.
Bring to review
A responsibility map covering context, tool permissions, handoffs and final decision ownership.
S01 sourceS02 source

Optional overlay

Reflexion-style feedback

Feedback from an attempt becomes textual memory that can guide a later attempt.

  1. Attempt and check
  2. Record useful feedback
  3. Bounded next attempt
Conceptual flow. Arrows show order. Repeat paths require explicit limits.
Useful when
An iterative read-only task has a useful external check and room for another model call.
Reconsider when
Feedback is unreliable, memory retains sensitive data, or retrying could repeat a write.
Bring to review
A feedback rubric, memory retention boundary and an explicit retry limit.
S05 source

How the decision is made

Sources checked

The selector applies a published rule order to the eight inputs. It does not call a model or estimate a probability of success. These are editorial design rules. The cited publications explain mechanisms and tradeoffs, not this exact decision tree.

  1. Reject every contradictory constraint before selecting.
  2. Single task: single model call.
  3. Known steps: fixed workflow, including configured branches.
  4. Separate specialist contexts: supervisor.
  5. Planned independent work: dependency-aware execution.
  6. Other planned work: plan and execute.
  7. Otherwise: ReAct.
  8. Add feedback only to eligible iterative, read-only work.

Read the status precisely

Candidate
A pattern fits this rule set. Validate it against representative tasks and failures.
Control gap
A candidate exists, but the declared write checks are incomplete.
Needs redesign
The inputs contradict one another. Every conflict remains in the export.

Primary sources and their limits

Open each record for provenance.

S01 · Building effective agents

Anthropic. Checked .

Pattern concepts only. The page flags changed tooling. No current SDK recommendation is inferred.

Read the primary source

Retained snapshot SHA-256: 6e137644bfee05b4fcfecc1db9843114db36e4821a110c9633880da2dc61a915

S02 · AI agent orchestration patterns

Microsoft Azure Architecture Center. Checked .

Descriptive architecture guidance, not comparative performance evidence.

Read the primary source

Retained snapshot SHA-256: a2e52539ff25ccf06cb42f1dda4ef71ac0c750010cadcb56bfd84a3ba33b3e8b

S03 · ReAct: Synergizing Reasoning and Acting in Language Models

Yao et al.. Checked .

Abstract-level mechanism evidence. Reported benchmark gains are not transferred to this selector.

Read the primary source

Retained snapshot SHA-256: a343bf95baa1521c0010ae6467049b6296bc38c91797eea28bd696acbb7b1606

S04 · An LLM Compiler for Parallel Function Calling

Kim et al.. Checked .

Dependency-aware execution concept. No speed or cost improvement is promised.

Read the primary source

Retained snapshot SHA-256: aea2c7d8ac159a23e9bb17ee8eb9f7da89ebeef244778e9e9853f98c6373deec

S05 · Reflexion: Language Agents with Verbal Reinforcement Learning

Shinn et al.. Checked .

Verbal feedback and episodic memory. This selector imposes stricter external-feedback and read-only conditions than the paper.

Read the primary source

Retained snapshot SHA-256: 23213e5703f133ec82acf956daf67ea1bd4ae3ea78b03b5580d906ed826b3589

S06 · Plan-and-Execute Agents

LangChain. Checked .

Planner/executor/replanning concepts, not a current API implementation contract.

Read the primary source

Retained snapshot SHA-256: ac3328daec0d87b6db5be72f5caa8b933e94868b9816fb79ac19fa97b37d2f46

S07 · AI Agent Architecture Patterns: A Decision Framework

Publisher architecture guide. Checked .

Publisher method evidence for scoped tools and write checks. Not independent validation or a measured outcome.

Contextual guide link

Retained snapshot SHA-256: 270803c194cd3418e16b8d953eaca200d2a606fd6d5826522a29d5e9d92f0ca9

Source snapshots document what was inspected. Their hashes identify retained text. They do not prove the claims true or guarantee that the linked page is unchanged. Full third-party text is not redistributed.

Portable decision record

Keep the inputs with the answer

JSON and Markdown exports contain all eight inputs, conflicts, review controls, alternatives, pattern definitions and source bindings. Identical inputs produce identical exports for this version.

First release: . No prior-version comparison is implied. Revisions receive a new catalog and rule version.

Rules / catalog
2026-09-18.1 / schema 1
Source JSON SHA-256
e97e4516e8cd9ef9d2492e04f7fc7ece4c04ef39ee9266ce04b194a6321f9374
Download full source JSONInspect the deterministic rule moduleDownload file integrity manifestSource repository and corrections