Agentic Loops

Reference entry · last updated 20260910

An agentic loop is a repeated execution cycle in which an AI agent uses observations from its environment to choose subsequent actions toward a task.[1]

Previous version, captured 20260910

First principles and definitions

In Anthropic’s distinction, workflows follow predefined code paths; agents dynamically direct their tool use. A workflow can contain a retry loop without delegating action selection to a model.[1] Feedback can change the next action, but does not guarantee progress or a correct result.

The agent harness executes tool calls and manages the surrounding runtime. Trivedy includes orchestration logic within the harness.[2] Prompt engineering shapes instructions; context engineering determines the information available during execution.

Task and current observations

↓

Model chooses an action

↓

Harness executes an allowed tool

↓

Inspect result and stopping conditions

↺ Continue with new observations

or ↓ Stop

Completion · Budget or time exhausted · Cancellation
Unrecoverable failure · Human escalation

Illustrative control flow. The host can interrupt execution independently of the model.

Worked example

This illustrative research task uses read-only tools: find the publication year of a named paper and return its source.

The agent searches the title and receives several candidate records.
It opens the matching publisher record to check the title and year.
If the record describes a different edition, that observation changes its next search.
It returns the verified year and citation, or reports the unresolved question when its search limit is reached.

The changed query demonstrates feedback-driven action selection. The example does not require a filesystem write or another external mutation.

Architecture patterns

These patterns describe different responsibilities, not ranked autonomy levels.

ReAct interleaves reasoning and actions within a task. Human approval is a separate deployment policy.[3]
Reflexion adds verbal reflection retained in episodic memory for later attempts. Its mechanism does not rewrite tools or update model weights.[4]
An orchestrator can assign subtasks to workers and combine their results. An evaluator-optimizer workflow instead revises an output using evaluator feedback. Anthropic describes both patterns separately.[1]

Weng describes planning, memory, and tool use as components of an LLM-powered agent.[5]

Design considerations

Anthropic recommends starting with the simplest effective system and weighing added cost and latency against performance.[1] Task frequency affects that tradeoff; it is not a definition of an agentic loop.

Feedback: Tests, tool results, human judgments, or model evaluations can guide another attempt. Evaluation must match the task; a passing check covers only what it measures.
Tools: Available operations can be read-only. Mutation permissions should match the task and be enforced by the host.
Bounds: Set limits on turns, time, and expenditure, with a defined outcome when a limit is reached.

The following stopping and recovery rules are illustrative engineering policies, not requirements of ReAct or Reflexion.

Failure modes and stopping

Repeated failure: inspect the error, revise the plan or inputs, and cap retries. Escalate or stop when the failure remains unresolved.
Lost task context: retain the goal and relevant evidence within the available context window. Record unresolved assumptions explicitly.
Ambiguous tool outcomes: check whether a mutation took effect before retrying it, to avoid duplicate side effects.[6]
Untrusted instructions: treat retrieved content as evidence. Prompt injection defenses and permission checks belong in the surrounding system; a prompt alone is not containment.

A fail-closed policy can stop restricted actions when authorization cannot be established. Cancellation, hard limits, and tool permissions should remain enforceable even when the model proposes continuing.

References

Anthropic. “Building effective agents.” 2024.
Trivedy, Vivek. “The Anatomy of an Agent Harness.” LangChain, 2026.
Yao, Shunyu, et al. “ReAct: Synergizing Reasoning and Acting in Language Models.” ICLR, 2023. Author project page.
Shinn, Noah, et al. “Reflexion: Language Agents with Verbal Reinforcement Learning.” NeurIPS, 2023. Full text.
Weng, Lilian. “LLM Powered Autonomous Agents.” 2023.
Amazon Web Services. “Making retries safe with idempotent APIs.” Amazon Builders’ Library.