Ⅱ Ⅱ~/IdeaProjects/ultimate

▲ 1 MCP server needs authentication · run /mcp

› I know the concept of agent loops is popular now. Explain what these agent loops are, how to benefit and set them up.

Ran 1 shell command

● Short answer up front: an “agent loop” is model → tool call → observe result → decide again, repeated until a stop condition.

The inner loop (what an agent already is)

prompt ──→ LLM ──→ tool call ──→ real-world result ──→ back into context ──→ LLM ──→ …

The thing that makes it work rather than just spin is the feedback signal. An agent editing code with tests in the loop converges.

The outer loop (what people are excited about)

Re-running the agent, unattended, on a trigger. Interval loops, scheduled cloud agents, and fan-out subagents each add another layer of text and commands.

〉 Ask another question…

PhpStorm Hackathon · The proposal

Move beyond the prompt box.

Three product ideas turn the coding agent from a text conversation into an IDE-native platform.

Skill Factory

Build a skill for the project currently open: discover a repeated problem, design the behavior, try it, and compare the result.

teach your IDE agent

Agent Apps

Embed purpose-built websites in the editor and interact with agents through structured screens instead of walls of chat.

use agents visually

Agent HTTP API

Expose IDE agents to localhost, websites, and IDE actions so any surface can start real agent work.

trigger agents anywhere

Idea 01 · Skill Factory

Teach the agent from inside your project.

Skill Factory is a PhpStorm feature. It works with the codebase, configuration, tests, and agent sessions already open in the IDE.

1ExploreFind friction in this project

2DefineChoose one better behavior

3DraftWrite with IDE context

4TryUse it on real work

5CompareWith skill vs without

6UseEnable and keep improving

1.8→90.2%correct with php-runtime-tools

53%lower cost

66%faster

The method already produced a high-impact PhpStorm skill. The product makes that capability available to every PhpStorm user and codebase.

Step 01 · Explore

Find friction in the project you are working on.

Skill Factory reads the open codebase and agent history with you. Start from a correction you keep making—not a generic idea.

A

Read project context

Framework, dependencies, run configurations, inspections, tests, and local conventions.

B

Review agent sessions

Spot repeated mistakes, missed IDE knowledge, and instructions you keep typing again.

C

Pick one problem

You choose the behavior worth fixing. The factory keeps the scope specific to this codebase.

OutputA concrete “the agent does X here, but should do Y” problem from the open project.

Step 02 · Define

Describe what “better” means in this codebase.

PhpStorm and the developer define the trigger, the desired outcome, and which project knowledge the agent should use.

01

When should it activate?

Name the task, file type, framework, or repeated request that should bring the skill in.

02

What should change?

Turn the developer’s correction into a short rule with a visible, checkable result.

03

What can the IDE supply?

Select project model, inspections, navigation, run configurations, or other PhpStorm tools.

OutputA focused skill brief grounded in the current project and approved by the developer.

Step 03 · Draft

Build the skill beside the code it will help with.

The agent drafts the skill while PhpStorm supplies real symbols, project settings, tool outputs, and conventions from the open workspace.

✎

Generate the first draft

Concise instructions, a clear trigger, and the specific PhpStorm capabilities to use.

↔

Preview as you edit

See which requests would activate the skill and inspect the context it adds.

✓

Review together

The developer changes wording, narrows scope, and removes assumptions that do not fit the project.

OutputA project-aware skill draft ready to try immediately in the open workspace.

Step 04 · Try

Use the draft on work that matters now.

Select a task in the open project and let the agent work with the skill. PhpStorm observes the result through the project’s own checks.

1

Choose a representative task

A real refactor, test failure, framework command, or project-specific workflow.

▶

Let the agent use the skill

Watch when it activates, which IDE context it requests, and what it changes.

✓

Check the project outcome

Use tests, inspections, build output, diffs, or a developer review already available here.

OutputA real attempt, its project checks, and clear feedback on where the draft helped or failed.

Step 05 · Compare

See whether the skill actually improves this project.

Repeat the same project task with and without the skill. Keep the model, code, settings, and acceptance checks the same.

OFF

Without the skill

Capture the agent’s normal approach and the result from the project’s own checks.

ON

With the skill

Run the same scenario with the draft active and record the same signals.

Δ

Show the difference

Correctness first, then time, cost, tool use, changed files, and developer intervention.

OutputA simple project-level comparison the developer can trust before enabling the skill.

Step 06 · Use & improve

Make it part of how this project works.

Enable the skill for yourself or the project team. Keep its result visible and reopen it whenever the codebase or workflow changes.

ME

Enable for me

Keep the skill local when it captures a personal workflow or preference.

PRJ

Share with the project

Store a team-approved skill with the codebase so every agent follows the same guidance.

↺

Improve from new sessions

When the skill misses, bring that session back into the factory and compare the revision.

OutputA useful skill attached to a real codebase—with its purpose and evidence still visible.

Idea 02 · Agent Apps

Give every agent workflow a real interface.

Embed websites like the Skills control room directly in the editor. The page explains state, collects decisions, and sends structured work to agents.

01Purpose-built, not genericA decision queue looks like decisions. An evaluation looks like results—not chat bubbles.

02Two-way interactionForms trigger agents; the agent updates the page when the work changes.

03IDE context stays attachedThe app runs beside the project, tools, sessions, and code it controls.

What needs your decision

The agent did the reading. You make the ruling.

How should the negative test set be built?Two supported options, cost and risk already compared.

Review decision

Run the next evaluation stageInputs, constraints, and expected output are shown before launch.

Open launch form

PhpStorm runtime skillBundled · 33 tasks · evidence attached

Idea 03 · Agent HTTP API

The IDE agent becomes a local platform.

Expose opted-in agent sessions as HTTP endpoints. Websites, localhost tools, and IDE actions can start work without pretending to be a user typing a prompt.

Any trusted surface

embedded Agent App

localhost website

IDE toolbar action

plugin or script

→

PhpStorm Agent Gateway

POST /agents/{session}/messages

POST /actions/{action}

GET /runs/{id}/events

permission + project scope

→

IDE agent runtime

project context

IDE + MCP tools

approval boundaries

Review selectionBuild evalExplain failure

The combined vision

PhpStorm becomes where agents are built, used, and connected.

The prompt remains available. It is no longer the only interface.

01 · BUILD

Skill Factory

Let PhpStorm users teach the agent with the project, tasks, and checks already open.

02 · USE

Agent Apps

Let structured, interactive websites live inside the IDE and collaborate with agents.

03 · CONNECT

Agent HTTP API

Make IDE agents safely reachable from buttons, plugins, scripts, and local web products.

Hackathon goalDemonstrate all three as one coherent IDE-native experience.