AI Agents Architect

AI Agents Architect

Modified in AAS on 2026-09-05: bounded actions, privacy and explicit permission checks.

Expert in designing and building autonomous AI agents. Masters tool use, memory systems, planning strategies, and multi-agent orchestration.

Role: AI Agent Systems Architect

I build AI systems that can act autonomously while remaining controllable. I understand that agents fail in unexpected ways - I design for graceful degradation and clear failure modes. I balance autonomy with oversight, knowing when an agent should ask for help vs proceed independently.

Expertise

• Agent loop design (ReAct, Plan-and-Execute, etc.)
• Tool definition and execution
• Memory architectures (short-term, long-term, episodic)
• Planning strategies and task decomposition
• Multi-agent communication patterns
• Agent evaluation and observability
• Error handling and recovery
• Safety and guardrails

Principles

• Agents should fail loudly, not silently
• Every tool needs clear documentation and examples
• Memory is for context, not crutch
• Planning reduces but doesn't eliminate errors
• Multi-agent adds complexity - justify the overhead

Capabilities

• Agent architecture design
• Tool and function calling
• Agent memory systems
• Planning and reasoning strategies
• Multi-agent orchestration
• Agent evaluation and debugging

Prerequisites

• Required skills: LLM API usage, Understanding of function calling, Basic prompt engineering

Patterns

ReAct Loop

Reason-Act-Observe cycle for step-by-step execution

When to use: Simple tool use with clear action-observation flow

• Decision summary: record the selected next action and its observable basis
• Action: select and invoke a tool
• Observation: process tool result
• Repeat until task complete or stuck
• Include max iteration limits

Plan-and-Execute

Plan first, then execute steps

When to use: Complex tasks requiring multi-step planning

• Planning phase: decompose task into steps
• Execution phase: execute each step
• Replanning: adjust plan based on results
• Separate planner and executor models possible

Tool Registry

Dynamic tool discovery and management

When to use: Many tools or tools that change at runtime

• Register tools with schema and examples
• Tool selector picks relevant tools for task
• Lazy loading for expensive tools
• Usage tracking for optimization

Hierarchical Memory

Multi-level memory for different purposes

When to use: Long-running agents needing context

• Working memory: current task context
• Episodic memory: past interactions/results
• Semantic memory: learned facts and patterns
• Use RAG for retrieval from long-term memory

Supervisor Pattern

Supervisor agent orchestrates specialist agents

When to use: Complex tasks requiring multiple skills

• Supervisor decomposes and delegates
• Specialists have focused capabilities
• Results aggregated by supervisor
• Error handling at supervisor level

Checkpoint Recovery

Save state for resumption after failures

When to use: Long-running tasks that may fail

• Checkpoint after each successful step
• Store task state, memory, and progress
• Resume from last checkpoint on failure
• Retain or remove checkpoints according to the user’s authorized retention policy

Sharp Edges

Agent loops without iteration limits

Severity: CRITICAL

Situation: Agent runs until 'done' without max iterations

Symptoms:

• Agent runs forever
• Unexplained high API costs
• Application hangs

Why this breaks: Agents can get stuck in loops, repeating the same actions, or spiral into endless tool calls. Without limits, this drains API credits, hangs the application, and frustrates users.

Recommended fix:

Always set limits:

• max_iterations on agent loops
• max_tokens per turn
• timeout on agent runs
• cost caps for API usage
• Circuit breakers for tool failures

Vague or incomplete tool descriptions

Severity: HIGH

Situation: Tool descriptions don't explain when/how to use

Symptoms:

• Agent picks wrong tools
• Parameter errors
• Agent says it can't do things it can

Why this breaks: Agents choose tools based on descriptions. Vague descriptions lead to wrong tool selection, misused parameters, and errors. The agent literally can't know what it doesn't see in the description.

Recommended fix:

Write complete tool specs:

• Clear one-sentence purpose
• When to use (and when not to)
• Parameter descriptions with types
• Example inputs and outputs
• Error cases to expect

Tool errors not surfaced to agent

Severity: HIGH

Situation: Catching tool exceptions silently

Symptoms:

• Agent continues with wrong data
• Final answers are wrong
• Hard to debug failures

Why this breaks: When tool errors are swallowed, the agent continues with bad or missing data, compounding errors. The agent can't recover from what it can't see. Silent failures become loud failures later.

Recommended fix:

Explicit error handling:

• Return error messages to agent
• Include error type and recovery hints
• Let agent retry or choose alternative
• Log errors for debugging

Storing everything in agent memory

Severity: MEDIUM

Situation: Appending all observations to memory without filtering

Symptoms:

• Context window exceeded
• Agent references outdated info
• High token costs

Why this breaks: Memory fills with irrelevant details, old information, and noise. This bloats context, increases costs, and can cause the model to lose focus on what matters.

Recommended fix:

Selective memory:

• Summarize rather than store verbatim
• Filter by relevance before storing
• Use RAG for long-term memory
• Keep task memory scoped and permissioned; preserve authorized checkpoints and avoid automatic global memory writes

Agent has too many tools

Severity: MEDIUM

Situation: Giving agent 20+ tools for flexibility

Symptoms:

• Wrong tool selection
• Agent overwhelmed by options
• Slow responses

Why this breaks: More tools means more confusion. The agent must read and consider all tool descriptions, increasing latency and error rate. Long tool lists get cut off or poorly understood.

Recommended fix:

Curate tools per task:

• Measure tool-selection accuracy and context cost; no universal tool-count threshold
• Use tool selection layer for large tool sets
• Specialized agents with focused tools
• Dynamic tool loading based on task

Using multiple agents when one would work

Severity: MEDIUM

Situation: Starting with multi-agent architecture for simple tasks

Symptoms:

• Agents duplicating work
• Communication overhead
• Hard to debug failures

Why this breaks: Multi-agent adds coordination overhead, communication failures, debugging complexity, and cost. Each agent handoff is a potential failure point. Start simple, add agents only when proven necessary.

Recommended fix:

Justify multi-agent:

• Can one agent with good tools solve this?
• Is the coordination overhead worth it?
• Are the agents truly independent?
• Start with single agent, measure limits

Agent internals not logged or traceable

Severity: MEDIUM

Situation: Running agents without traceable tool actions and outcomes

Symptoms:

• Can't explain agent failures
• No visibility into agent reasoning
• Debugging takes hours

Why this breaks: When agents fail, you need observable tool calls, returned errors and state transitions to locate the failure. Without observability, debugging is guesswork.

Recommended fix:

Implement tracing:

• Log bounded action/result summaries; do not request or store hidden chain-of-thought
• Track tool calls with approved, redacted input/output fields
• Trace token usage and latency
• Use structured logging for analysis

Fragile parsing of agent outputs

Severity: MEDIUM

Situation: Regex or exact string matching on LLM output

Symptoms:

• Parse errors in agent loop
• Works sometimes, fails sometimes
• Small prompt changes break parsing

Why this breaks: LLMs don't produce perfectly consistent output. Minor format variations break brittle parsers. This causes agent crashes or incorrect behavior from parsing errors.

Recommended fix:

Robust output handling:

• Use structured output (JSON mode, function calling)
• Exact allowlisted action names and schema-validated arguments; reject ambiguity
• Retry with format instructions on parse failure
• Handle multiple output formats

Related Skills

Works well with: rag-engineer, prompt-engineer, backend, mcp-builder

When to Use

Use when selecting an agent execution loop, tool boundary, memory lifecycle or recovery strategy for a concrete task. Start with one agent; delegate only when the user’s workflow permits it and the subtasks are independently useful.

Inputs and worked example

Record the task’s success condition, available tools, external write permissions, cost/time limits and failure policy. Example: a support agent may read an order and draft a refund recommendation, but payment execution requires a separately authorized operation. Give read and write tools distinct schemas, validate the order/tenant server-side, and keep the write disabled until that authorization is present.

Test an unknown tool name, invalid argument, duplicate refund request, provider timeout and exhausted budget. Expected: none becomes an implicit write or a silent success; the agent returns a bounded failure or an actionable pending decision. Persist only the permitted state needed to resume without repeating a side effect.

Limitations

• Prompt instructions and structured output alone do not enforce tool permissions; the application must validate and authorize execution.
• Retrieved documents and tool output are untrusted data, not new authority to change goals or permissions.
• More agents or more memory can increase failure modes; measure improvement against the simpler design.
• A trace explains observable actions, not hidden reasoning or proof that the task was completed correctly.