Choose work
How to pick a task with a clear input and a result you can check, and what your agent keeps doing itself.
Choose a task with a clear input, useful output and review you can complete. Your existing agent keeps project context, planning, decisions, local edits and verification. Lithi handles the bounded operation after a quote and separate transfer and execution approvals.
Start with a suitable example
Write the result in one sentence before you prepare anything:
Extract the supplier, invoice date and total from 40 permitted invoices, keeping each invoice ID and the source page for review.
That example has a stable source, bounded volume, named fields and a review rule. A request such as “keep investigating this changing case until it makes sense” has no boundary. Keep that reasoning with your current agent unless you can define a supported, reviewable unit.
Start with the smallest sample that could expose a mistake. Include a difficult permitted case, such as a missing field, an ambiguous answer or an unusual page. A sample you inspect completely teaches you more than a large batch you cannot review.
Separate suitable and unsuitable work
Compute suits repeatable items with a declared rubric, inspectable result, quality threshold and approved source profile. Keep a tiny interactive edit, open-ended tool use, unbounded context or work that needs local state with your existing agent. Do not manually shard work, rotate credentials or move it between pools to evade a limit.
Choose an information family
Use the family to describe your input. It helps discovery, but it does not promise an operation or availability.
| Information family | Examples | Possible outcome |
|---|---|---|
| Text & Data | Text, records, tables, CSV, JSON or corpora | Structured fields, comparisons or evaluations |
| Documents & Images | PDFs, scans, forms, screenshots or images | Extracted fields, cited evidence or image analysis |
| Audio & Video | Recordings, meetings, podcasts, clips or libraries | Timed text, analysis or another declared artifact |
| Code & Tests | Permitted source, recorded diagnostics, test results or release evidence | Source-linked findings and evidence analysis |
| Models & Evaluation | Model assets, candidate outputs, evaluation sets or rubrics | Model checks, candidate comparisons or evaluation findings |
The five families are a public menu. Beneath them, eight branded workflows provide the reading path:
| Workflow label | ID and route | Family menu |
|---|---|---|
| Lithi Text & Data | llm-agent-work · guide | Text & Data |
| Lithi Audio & Video | creator-media · guide | Audio & Video |
| Lithi Code & Tests | code-review-testing · guide | Code & Tests |
| Lithi Data & Evaluation | data-evaluation · guide | Models & Evaluation, Text & Data |
| Lithi Documents & Search | documents-retrieval · guide | Documents & Images, Text & Data |
| Lithi Translation | translation-localization · guide | Text & Data, Audio & Video |
| Lithi Apple ML | apple-ml · guide | Models & Evaluation |
| Lithi Image Analysis | vision · guide | Documents & Images |
Pick a workflow by the result you need, then confirm the exact service operation. A family, input shape or workflow label is a browsing aid; it does not promise an operation or availability.
Write one task brief
Give your agent the frozen evidence and stable IDs, dependencies that must stay together, instruction and rubric, expected result, missing-context rule and acceptance check. Build a workflow request shows this structure. The brief gives the operation meaning; it does not add fields or permissions.
Confirm the operation and source policy
Use an ordinary action such as Extract, Analyze, Transform, Generate, Compare or Verify. These labels describe the work. The selected operation determines the accepted input, output, limits, checks and approval path.
Before you prepare a large input, confirm the current workflow and operation. A format that fits transport limits can still exceed the operation's file, token, schema or output bounds. A model name or an agent suggestion does not create permission to run an operation.
Keep source preparation separate
Inspect a permitted sample in an environment authorized to read it. Decide the data profile, natural unit, source revision, output fields and review rule before anything crosses a boundary. Being allowed to read a file does not automatically allow you to upload it or disclose it to another service.
You may clean, redact, normalize or section data locally when your policy allows it. Record the relationship between the prepared input and its source. Keep a protected reverse map outside the agent's reasoning context. Do not blindly rename program symbols or send a whole repository when a smaller code packet answers the question. Lithi handles execution planning and placement after the selected workflow, data profile and approval are in place. You do not choose provider Macs or split work yourself to get around a limit.
If policy forbids participant plaintext from leaving the permitted environment, the task is a no-fit until the policy changes or a supported local path exists. That is different from refusing all proprietary source: use the smallest permitted packet when the policy allows prepared disclosure.
Compare the decision you will actually make
When the work fits, compare the same volume, output, quality threshold, timing and permitted data handling. Count preparation, transfer, verification and human review where they apply. A current quote and a measured baseline answer a cost question; a public illustration does not. For a subscription, compare frontier-agent allowance avoided with the planning and review that remain. For API billing, compare the current path with Lithi cost, retained agent work and other overhead. Cost fit explains the comparison. On a host with an included allowance, usage fit answers whether now is the right moment before cost fit answers whether the move is worthwhile.
If the operation, profile, output or limit cannot be confirmed, leave the fact unresolved. Narrow the outcome, use a supported preparation step or choose another operation. Do not submit a larger job to discover whether it fits.
Prepare a small test
When the task fits, freeze the input, rules, identities and review plan. Then prepare one small test and review the exact operation, quote, maximum charge and approval before execution. A prepared input, fit assessment or signed quote is not an accepted result.