Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
An AI agent is a system that repeatedly thinks, acts through tools, observes what happened, and uses the new information to choose what to do next. That interaction—not a long prompt or an impressive-sounding model—is what makes an agent different from a one-shot answer or a fixed workflow. The five waves of agent engineering show how the work expands from prompting to designing the full execution system around a model.
What an AI agent is actually doing
An agent operates in a loop: it receives context, decides on an action, acts on an environment, observes the result, and decides again. The observation matters because it can reveal facts unavailable when the first action was chosen: a compiler error, a changed file, a failed tool call, or another result from the environment.
This is why interaction cannot always be replaced with a longer static answer. A prompt can describe likely outcomes, but it cannot know the actual output of a command that has not run. As DogeKing puts it in the October 2, 2026 DEV Community article, “An Agent’s action trajectory cannot be reduced to one longer static answer.”
Agent, workflow, or one API call?
A single stateless model call generates a response from the context it receives; it does not independently act on an environment and adapt to the result. A fixed workflow can execute multiple steps, but its path is predetermined. An interactive agent is distinguished by whether unexpected results can change its next action.
#1 Best Overall
A useful test is to ask what happens when a tool returns something unexpected. If the system can use that observation to correct course, it has an interactive loop. If the sequence cannot change, it is closer to a fill-in-the-blank prompt or fixed program than an autonomous agent. This distinction also helps avoid inflated claims: a model saying that tests passed is not evidence that tests were run, and a long chain of dependent steps can propagate errors rather than demonstrate sound judgment.
How the ReAct loop works in CodeSmith
The CodeSmith example is based on source version v0.5.0, commit 3a74c82f. Its DefaultAgentExecutor::run_inner is described as implementing a ReAct-style loop: reason about the task, act with tools, then incorporate observations before continuing.
- The executor assembles a message request and streams a model response.
- It collects any tool calls requested by the model and executes them.
- It adds the tool results to the conversation history, represented under the user role.
- It sends the updated history back to the model and repeats while the model continues requesting tools.
- It stops when the model no longer requests tools or an execution exit condition is reached.
If a requested tool does not exist, CodeSmith returns a NotAvailable result to the model. That gives the model a chance to respond to the failure instead of making the missing tool an immediate crash.
Stopping is part of the loop
The executor’s stop enum has four exits: NoToolCalls, MaxSteps, Error(String), and Interrupted. The article reports that max_steps defaults to 50. A step limit can bound an otherwise continuing interaction; the other exits distinguish a completed tool-use turn from a failure or interruption.
Why context and the tool interface affect capability
The model can only use the information and actions made available to it. In the article’s context ablation, four components play different roles: tool definitions make actions available, tool results provide feedback, reasoning records why an action was chosen, and message history helps avoid redundant operations and repeated mistakes. Removing context can still leave a fluent response, but fluency alone does not show that the task was completed.
The interface matters too. The article points to SWE-agent as an example of how the same foundation model can behave differently with a plain shell versus a purpose-designed Agent-Computer Interface. The way files are presented, how edits are issued, and what error messages say all shape what the model can do with its tools. Capability is therefore not just a property of model weights; it also depends on the harness and the environment the model must navigate.
Autonomy is a spectrum, not a yes-or-no label
The CodeSmith article places agent autonomy on five levels. The levels describe who controls the action sequence and whether the system can question or revise the task, not a universal score for agent quality.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors| Level | What the system controls |
|---|---|
| 1 | The developer specifies every action. |
| 2 | The model selects actions from available tools. |
| 3 | The model revises its plan after surprises. |
| 4 | The model proposes and decomposes subgoals. |
| 5 | The model examines the task and its evaluation criteria. |
CodeSmith is described as between levels 2 and 3: it selects tools and can be pushed toward verification and replanning. The distinction is meaningful. Choosing a tool from a menu is less autonomous than detecting that a result invalidates the plan and forming a new one.
How agent loops handle memory and replanning
Not every agent remembers in the same way. The loop and memory families compared in the article differ in what they retain, what they retrieve, and what triggers another round.
| Approach | What is saved | What is read | What triggers the next round |
|---|---|---|---|
| ReAct | No extra cross-step memory | Current conversation and tool results | The model continues the think–act–observe loop |
| Reflexion | A reflection after failure | Stored reflection as guidance | A failed attempt informs another try |
| LATS | Alternative paths in a search tree | Candidate paths, with backtracking | Search explores or revisits branches |
| Voyager | Successful skills | Previously learned skills | A task calls for a skill or a new one to be learned |
| MemGPT | Layered memory | Relevant memory pages as needed | Paging brings information into context |
These are different strategies, not interchangeable labels for “having memory.” A system that retains a failure reflection is solving a different problem from one that searches multiple candidate paths or pages information from layered storage. The right design depends on what must persist and how future actions should use it.
When multiple agents help—and when they do not
Adding agents does not guarantee better results. Delegation is most useful when the work can be divided into sufficiently independent subtasks and the coordinating system can manage dependencies between them. If subtasks rely on one another, coordination must make those relationships explicit; otherwise parallel work can produce incompatible or duplicated results.
Recommended Free Tools
A delegation contract should define the objective, permitted tools, forbidden actions, resource ceilings, abort conditions, output format, responsibility boundaries, and how the task can be renegotiated. Clear boundaries make it easier to know what a delegated agent may do and what the coordinator still owns.
Rank #4
Agreement among agents is not automatically independent confirmation. Agents drawing on the same source or assumptions can converge on the same error, and debate can amplify anchoring rather than correct it. Team size therefore has coordination costs and gains that depend on task structure, not a simple “more is better” relationship.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The five waves of agent engineering
The article presents five nested waves. Each expands the engineering surface rather than making the earlier work obsolete.
1. Prompt engineering
Prompt engineering optimizes the natural-language instructions given to the model. It shapes how the model interprets the task, but cannot supply observations that only an action can produce.
Free tools Windows power users keep installed
One-click scans. No signup required.
2. Context engineering
Context engineering manages everything the model can see: instructions, conversation history, tool definitions, tool results, and other relevant information. It determines what the model can use to choose its next step.
Best Value
3. Harness engineering
Harness engineering designs the surrounding system: tools, constraints, verification, feedback, and recovery. It governs how model choices become actions and how results return to the model.
4. Loop engineering
Loop engineering sustains operation across turns. It addresses when to act, when to verify, how to respond to surprises, and when to stop rather than continue indefinitely.
5. Graph engineering
Graph engineering organizes agent loops alongside deterministic programs and human approvals in an execution graph. Not every step needs to be delegated to a model; a graph can place model judgment where useful and retain explicit program logic or human review where needed.
The larger shift is from treating the model as the whole product to treating it as one component in a system of context, interfaces, checks, loops, and coordination. The waves remain nested: good prompts still matter inside a well-designed context and harness.
What the reported benchmark change does—and does not—show
DogeKing’s October 2, 2026 article reports that LangChain’s Terminal Bench 2.0 score rose from 52.8% to 66.5% after harness changes including automatic execution checks, repetitive-loop detection, and strategy refinement; the article says the model was not swapped. This is a reported example supporting the importance of harness design, not a general guarantee that those changes will produce the same gain in other systems or evaluations.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
