Methodology · Agent Constructionemergingverified

SPAR Agent Loop Design

also known as Sense-Plan-Act-Reflect, SPAR framework

Applies to: agentcoding-agentbrowser-agentvoice-agent

Tags: agent-loopsparsense-plan-act-reflectagent-anatomy

Build the agent's core as a loop with four named phases: Sense, Plan, Act, Reflect. Sense takes in signals from the environment and the task input. Plan works out what the agent intends to do next. Act runs the tool calls and makes the changes. Reflect checks the outcome against the plan and updates what the agent believes. Make each phase a real, named step you can point to in code and see in traces. Then you can tell which phase is failing instead of staring at a black box.

Methodology process overview

stateDiagram-v2 [*] --> Sense Sense --> Plan: observation ready Plan --> Act: plan emitted Act --> Reflect: outcome observed Reflect --> Sense: continue Reflect --> [*]: stop condition met Reflect --> Plan: replan needed note right of Act: only side-effects here

Intent. Give every agent the same four named phases, Sense, Plan, Act, and Reflect, so behaviour, traces, and failures line up with a phase instead of hiding in one murky loop.

When to apply. Use this when you design the main loop of any real agent: a coding agent, a research agent, a browser agent, or an ops agent. It helps most when a team is reaching for ReAct or a free-running loop with no clear control structure. Don't apply it for a single-shot prompt or a fixed pipeline with no perception and no reflection. There is no loop to design.

Example scenario

An infra team at a B2B SaaS company is rebuilding their on-call triage agent. The old version was one big ReAct prompt. It 'worked' but was impossible to trace when it misrouted an incident. They adopted SPAR. Inputs: a PagerDuty webhook stream as the environment, the goal 'route incident to the right responder with context', and an action set of runbook lookup, log query, Slack post, and PagerDuty re-assign. Build order: the Sense phase pulls the alert payload, the last 15 minutes of related logs, and any open incidents on the affected service. The Plan phase produces a typed plan ('lookup runbook X, then post summary, then re-assign to team Y') using chain-of-thought. The Act phase is the only one that can post or re-assign. Sense and Plan stay read-only. The Reflect phase checks the route against a rule of thumb: was the responder on-call, did they ack within 5 minutes? It then updates a belief about runbook accuracy. Results: traces now group cleanly by phase. When the agent misroutes, the team sees right away which phase failed. Two weeks in, they found Reflect was rubber-stamping bad routes because it shared the Plan prompt. They swapped in a different model for Reflect and the false-route rate dropped from 14% to 4%. The step budget caught a Plan-to-Reflect oscillation that would have looped forever in the old design.

Inputs

Environment interface — The channels the agent reads from: APIs, user turns, the file system, and other observers.
Task goal — The goal the agent is chasing, written so the Plan phase can break it into steps.
Action repertoire — The tools, effectors, or sub-agents the Act phase can use.

Outputs

Four-phase agent loop — Clear Sense, Plan, Act, and Reflect steps joined by typed transitions. You can test and watch each step on its own.
Phase trace schema — Structured trace events from each phase that your tooling can group, search, and replay.

Steps (6)

Build the Sense phase
Decide what the agent observes each tick: user input, tool outputs, memory recall, and environment state. Sense gathers the situation, it does not judge it. Surface anything unclear instead of papering over it.
usesAugmented LLM Scratchpad
Build the Plan phase
Turn the sensed situation into an intended next move or plan. Use chain-of-thought, tree-of-thoughts, or plan-and-execute when the task calls for it. The output is a plan, not a free monologue.
usesChain of Thought Plan-and-Execute Tree of Thoughts
Build the Act phase
Run the plan through tool calls, sub-agents, or outside effectors. Act is the only phase that may change anything. Sense, Plan, and Reflect stay read-only.
usesTool Use ReAct
Build the Reflect phase
Check the outcome of Act against what the plan intended. Update beliefs, note lessons, and decide whether to loop, replan, escalate, or stop. Reflect is what makes the agent more than a script.
usesReflection Self-Refine
Wire phase transitions and termination
Define the typed moves between phases: Sense to Plan to Act to Reflect, then back to Sense or stop. Add a step budget and clear stop conditions so the loop always ends.
usesStep Budget Composable Termination Conditions
Instrument each phase as a first-class trace event
Emit a structured trace event tagged with its phase. Your observability tools can then group, filter, and replay by phase, so failures stay easy to find.