--- title: "The agent" description: "How the agent works: the run loop, tools, approvals, and how to steer it" canonical_url: "https://hoplite.sh/docs/agent" markdown_url: "https://hoplite.sh/docs/agent.md" --- # The agent URL: /docs/agent LLM index: /llms.txt Description: How the agent works: the run loop, tools, approvals, and how to steer it Related: /docs/agent/models, /docs/agent/tools, /docs/threads # The agent Every Hoplite [thread](/docs/threads) is driven by an agent: a frontier model wired to a [sandbox](/docs/sandboxes) full of tools. It reads and edits your code, runs commands, drives a browser, and talks to GitHub — streaming everything it does into the thread. ## The loop Each message you send kicks off a **run** — one agent execution turn. During a run, the agent works through its [tool catalog](/docs/agent/tools): every call streams into the thread's activity timeline with a short description of what it's doing, and mid-run reasoning folds into the timeline so long runs stay scannable. Sensitive calls pause for your approval before they execute: edits to protected files or paths, shell commands, killing processes, sandbox restarts and downsizes, preview-checklist changes, and project-settings updates. Routine repository-local edits and reads, searches, browser actions, and status checks run without interruption so the agent keeps its momentum. See [Tools and permissions](/docs/agent/tools#which-tools-require-approval) for the full list. ## Web research and trust When staff enables it, a `web_search` tool searches current public-web documentation through Hoplite's Exa integration; each result includes its source URL. This is a platform capability, not an MCP server: the worker-owned Exa credential is never copied to project settings, repository files, or sandboxes. Search results are untrusted external content. The agent treats them as reference material, not instructions. ## Verifying its own work For a frontend change it can render, the agent captures and inspects fresh before-and-after screenshots, and shares focused video for meaningful interactions or transitions. It checks that screenshots and recordings do not expose sensitive account, tenant, credential, or other private data before sharing them. If evidence cannot be captured, redacted, or shared safely, the agent reports that blocker instead of silently omitting verification. ## Personalize agent output Under **Settings → Agent personalization**, **Agent verbosity** controls how much detail the agent writes back: low, medium, or high. **Code comments** controls the comments it adds to authored code: off, brief, or verbose. The defaults are medium verbosity and brief comments, and repository conventions still take precedence when they require a different style. These are per-user settings enforced for every run you start. Automatic continuations such as PR autofix keep the output preferences of the thread owner whose work they continue. ## Personal memories Under **Settings → Agent personalization → Agent memories**, you can review, create, edit, and delete durable personal memories. They are available only to agents you start; unattended automation runs without an initiating user do not inherit anyone's private context. The agent can save or update a concise, stable cross-thread preference without interrupting you for per-write approval. After a successful run, Hoplite may also curate an explicitly stated stable preference from your opening message. It does not infer preferences from agent output, tool results, repository content, or external data, and it rejects secrets, credentials, sensitive personal data, raw conversation text, repository or project facts, commands, branches, tickets, and temporary task state. Memory changes apply to later runs. A run already in progress keeps the immutable memory snapshot it started with, even if you edit or delete a memory from settings while that run is active. ## Self-configuration and feedback The agent can help configure itself: `project_settings_get` shows how the project's scripts and settings resolve, and `project_settings_update` can repair script overrides directly. Automatic-QA policy changes through that tool still require your approval. See [Project scripts](/docs/sandboxes/scripts). If the agent hits a platform problem — a broken sandbox, a misbehaving tool — it files a structured report to the Hoplite team with its `report_platform_issue` tool, so failures get fixed at the source instead of being worked around silently. ## In this section - **[Models and speed](/docs/agent/models)** — the model lineup, reasoning effort, speed tiers, and credential routing - **[Instructions](/docs/agent/instructions)** — standing project context prepended to every run - **[Skills](/docs/agent/skills)** — reusable instruction bundles the agent loads on demand - **[Slash commands](/docs/agent/slash-commands)** — composer shortcuts like `/pr`, `/model`, and `/retry` - **[Tools and permissions](/docs/agent/tools)** — the full tool catalog, and which tools require your approval - **[MCP servers](/docs/agent/mcp)** — extend the tool catalog with Model Context Protocol servers ## Sitemap See the full [sitemap](/docs/sitemap.md) for all pages. Well-known sitemap: [/docs/.well-known/sitemap.md](/docs/.well-known/sitemap.md).