> ## Documentation Index
> Fetch the complete documentation index at: https://docs.deeda.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Start Here: Pick by Task

> I want to ___ → the runtime, method, and blocks to use. The ELI5 router.

Find your task, copy the route. Every route links to a complete example.
(Agents: this page + [Snippets](/harness/snippets) + the
[Workflow Schema](/harness/workflow-schema) is the full assembly kit.)

| I want to…                                    | Runtime + method                                                   | Add these blocks                        | Example                                                                                 |
| --------------------------------------------- | ------------------------------------------------------------------ | --------------------------------------- | --------------------------------------------------------------------------------------- |
| Fix a bug / implement a feature in a repo     | `anthropic-agent-sdk` + `sdk`                                      | tools, sandbox(edit), budget            | [Minimal single agent](/harness/cookbook#recipe-minimal-single-agent)                   |
| Same, but survive outages/rate limits         | + ordered fallback to `openai-codex-sdk`, `gemini-antigravity-cli` | selection, retry                        | [Resilient worker](/harness/cookbook#recipe-ordered-fallback-resilient-single-agent)    |
| Research a topic, then build, then review     | `gemini-genai-sdk` → `anthropic-agent-sdk` → `openai-codex-sdk`    | pair\_profiles                          | [Pair recipe](/harness/cookbook#recipe-pair-programming-researcher--engineer--reviewer) |
| Get an unbiased code review                   | `openai-codex-sdk` + `cli`, `isolate_context: true`                | tool\_policy(read-only)                 | [Pair recipe](/harness/cookbook#recipe-pair-programming-researcher--engineer--reviewer) |
| Gate completion on multi-vendor agreement     | any worker + review panel                                          | review\_consensus                       | [Review panel](/harness/cookbook#recipe-provider-diverse-review-panel)                  |
| Extract structured data (JSON)                | `openai-responses-api` + `api` or any SDK runtime                  | state\_machine + output\_schema         | [Structured Output](/concepts/structured-output)                                        |
| Cheap bulk classification/evals               | batch surfaces                                                     | stage that submits, stage that collects | [Cost, Batch & Cache](/concepts/cost-batch-cache#batching)                              |
| Multi-stage pipeline with typed handoffs      | any                                                                | state\_machine                          | [State machine](/harness/cookbook#recipe-state-machine-workflow)                        |
| Hard problem — buy more reasoning             | Anthropic thinking / OpenAI reasoning\_effort / Antigravity level  | agent.thinking                          | [Thinking](/concepts/thinking)                                                          |
| Do a web task with no API (forms, dashboards) | `openai-operator` + `api`                                          | budget, approval-friendly policy        | [Computer Use](/concepts/computer-use)                                                  |
| Voice panel of experts                        | `roster` with `voice`/`voice_name`                                 | roster                                  | [Voice & Realtime](/concepts/voice-and-realtime)                                        |
| Keep data on-device / offline                 | `local-llm` + `api`                                                | budget                                  | [Local provider](/harness/providers/local)                                              |
| Block specific tool calls deterministically   | Claude-native hooks (settings)                                     | tool\_policy as backstop                | [Hooks](/concepts/hooks)                                                                |
| Plug in external SaaS tools                   | participant with `mcp_server_config`                               | tools                                   | [Tools & MCP](/concepts/tools-and-mcp)                                                  |
| Cut spend on repeated context                 | —                                                                  | cache\_policy, Headroom                 | [Cost, Batch & Cache](/concepts/cost-batch-cache)                                       |

## The 60-second mental model

1. **Workflow markdown is the only control plane.** Everything an agent may
   do — runtime, model, tools, network, budget, review — is declared in one
   file. If a power isn't declared, it doesn't exist at runtime.
2. **Cadence is the governor.** It validates your file, picks the concrete
   route, enforces budgets, and only advances stages on typed events.
3. **States are honest.** Every capability is `implemented`,
   model-dependent, partial-compatible, policy-blocked, vendor-blocked, or
   unsupported — each provider page has a [decision-rule legend](/harness/providers/anthropic#how-to-read-this-page).
4. **Test cheap first.** Validate (`think_workflow_md_validate`), run with a
   small `budget`, read the session ledger, then scale.