Agentic software-development tools

Volatile information — last reviewed 30 July 2026.

Product capabilities, interfaces, access conditions, pricing, policies, and maintenance status can change quickly. Check the linked official documentation before using this page in a session or adopting a tool.

This page lists representative tools that can inspect a project, propose or make changes, and use development tools on a user's behalf. It deliberately avoids installation commands, model names, context limits, and detailed feature comparisons because those details have a particularly short useful life.

For the durable concepts, scientific risks, and review workflow, see agentic AI and scientific software. Apply the general tool-selection criteria before choosing a product.

Representative current tools

This is a starting set, not an exhaustive ranking.

Tool	Current positioning	Points to verify for a project
Codex CLI	OpenAI's terminal client for interactive and non-interactive agentic software-development work. Current documentation describes workspace-oriented operation, repository guidance, and sandbox and approval controls.	Check the current security and approval model, authentication, data policy, supported platform, and whether local or hosted operation fits the project.
Claude Code	Anthropic's coding agent for terminal, editor, desktop, and web workflows. It can inspect code, edit files, run commands, and integrate with development tools.	Check execution location, permission mode, account or provider requirements, organizational policy, and which project instructions and integrations will be enabled.
Gemini CLI	Google's terminal agent for understanding code, automating tasks, and building workflows with project context. Its documentation includes scripting, project instructions, tools, policies, and sandboxing.	Check authentication, network requirements, data terms, release channel, sandbox configuration, and compatibility with restricted or offline systems.
GitHub Copilot CLI	GitHub's terminal agent for interactive or programmatic work on code and GitHub-hosted workflows.	Check the required Copilot plan and organization policy, repository and GitHub permissions, approval settings, sandbox availability, and whether the project is hosted on or integrated with GitHub.
GitHub Copilot code review	AI-assisted review of pull-request changes. A review can be requested manually or configured to run automatically when pull requests are opened, and optionally for drafts or new pushes.	Check plan and organization policy, usage costs, data handling, ruleset configuration, review instructions, and excluded files. Its comments do not count as required approvals or block a merge, so retain human review and appropriate CI checks.
Aider	A terminal-oriented, Git-integrated AI pair-programming tool that can use several model providers. It sits toward the interactive-assistant end of the agentic spectrum but can edit files and run project checks.	Check provider choice and data path, API-key handling, automatic commit behaviour, test and lint integration, model compatibility, and whether the desired workflow needs more or less autonomy.

The products overlap, but their trust boundaries differ. A terminal interface does not imply that model inference or data processing happens locally. Similarly, an open-source client does not by itself determine where prompts, source code, or tool output are sent.

Pull-request code review is also a different operating mode from a coding agent that edits a working tree. The version-controlled change provides a bounded difference to inspect, and the pull-request event can trigger both review and continuous integration. The review interprets the change and suggests issues; continuous integration runs the project's declared checks. They provide complementary evidence.

Compare operating modes, not feature counts

For a representative project task, determine

which files the tool reads and which context it sends elsewhere;
whether commands run on the workstation, in a container, or in a hosted environment;
which actions require confirmation and how permissions are scoped;
whether network access and external integrations can be restricted;
whether changes are easy to inspect, attribute, and reverse;
whether repository instructions, tests, and static checks are used consistently;
whether interactive and unattended modes have different safeguards; and
which logs, transcripts, or audit records are retained and by whom.

Do not select a tool primarily because it advertises the largest number of features or the most capable model. A more constrained tool may be preferable when the task is narrow, the data is sensitive, or the execution environment is shared.

Scientific-computing considerations

Evaluate each candidate with a realistic scientific repository rather than a small generic coding exercise.

Scientific correctness

Check whether the workflow makes it easy to provide and run

known-answer and regression tests;
numerical tolerances and invariants;
data-validation and schema checks;
small reference cases;
documentation builds; and
compiler warnings, linters, and static analysis.

These integrations make review more efficient, but no product-specific "verification" feature should be interpreted as proof of scientific correctness.

Data and confidentiality

Establish which source files, prompts, test data, command output, and repository metadata may be processed by the provider. Check institutional policy, contracts, data classifications, intellectual-property constraints, and retention settings.

Do not assume that excluding files from version control also excludes them from an agent's context. Check the tool's own ignore, permission, and context rules.

HPC suitability

For work involving a cluster, determine whether the tool

can operate without direct internet access on compute nodes;
requires installation or authentication that the site does not permit;
understands that expensive work must go through the scheduler;
can be prevented from reading SSH keys, tokens, unrelated storage, and other users' data;
separates local edits from remote job submission and cancellation; and
records enough compiler, library, module, accelerator, scheduler, and placement context to interpret results.

It may be preferable to use an agent on a workstation or controlled development node, then submit reviewed code through the site's normal workflow. Convenience does not override site policy or allocation responsibility.

A small evaluation task

This is an optional follow-up activity rather than part of the standard four-hour session.

Evaluate candidates on a disposable branch or copy of a real project. Give each tool the same bounded task, such as

Explain one calculation, add a scientifically meaningful test for an existing invariant, run the documented checks, and summarize the changes and remaining limitations.

Compare

the correctness and relevance of the explanation;
whether the test could actually expose a defect;
the size and focus of the change;
respect for repository instructions and scope;
commands and permissions requested;
quality of the reported limitations;
ease of reviewing and reversing the result; and
setup, policy, usage, and maintenance costs.

Do not use confidential data or grant broad permissions merely for an evaluation.

Maintenance checklist for this page

Before a training session or periodic tool review:

open every official documentation link;
confirm that the product and named surface are still maintained;
check whether the description of local, hosted, interactive, and non-interactive operation remains accurate;
review current permission, sandbox, data, account, and platform documentation;
remove deprecated terminology, commands, model names, and access claims;
add or remove candidates only when that improves the range of relevant choices; and
update the review date and summarize material changes in version control.

Stable selection principles belong in choosing development tools or the conceptual agentic AI chapter, not on this volatile page.