Which workflows should an agent run first?
Look for recurring work with clear inputs and a reviewable result. Examples include gathering research, drafting a service page or documenting a procedure. Start with copies of approved information and keep the output separate from the originals.
- It repeats at least twice a month, so the setup cost pays back
- It runs on inputs you already own: your notes, your site, your export
- It ends in a draft, never in a sent message or a live deploy
- It fails loudly, in a file you can read, not quietly in a customer inbox
If a task involves a customer commitment or an important record, add an approval step before that change occurs.
Five tasks to consider for a first trial
Each row identifies the task, its output and the action reserved for a person. Use these as starting points and adapt them to your own approval rules.
| Workflow | What the agent hands back | Gate a person keeps |
|---|---|---|
| Market and review research | Cited brief, themes, open questions | Which finding becomes a decision |
| Website and landing updates | Page brief, metadata, QA evidence | Production deploy and DNS |
| Lead preparation | Fit summary and unsent draft per lead | Send, list source, offer claim |
| Ad and social sets | Variant matrix, platform spec check | Spend, publish, every claim |
| SOP capture and reporting | Draft SOP, exception map, summary | Any write into a source system |
The review sits before an action affects customers, money or shared business records. The agent can prepare the material so the reviewer has a concrete result to assess.
What does one run look like end to end?
Keep the instructions, source files and generated drafts in clearly named locations. Restrict writing to the draft folder so a failed trial does not overwrite your source material.
work/outreach/
.claude/skills/lead-prep/SKILL.md # the job, in writing
input/leads.csv # yours; the agent reads only
out/2026-09-03/ # the only path the run may write
# run it, then read out/ before a single message goes anywhere
claude -p "Run lead-prep on input/leads.csv. Write drafts to out/2026-09-03/. Do not send."Where these workflows break
Use the following checks when reviewing a run. They help distinguish a finished-looking draft from work that is actually ready to use.
- Stale inputs. The brief cites a price you changed in June and reads authoritative anyway. Check dates before you check reasoning.
- Quiet scope creep. A skill written for one page starts rewriting the site. Read the diff, not the summary of the diff.
- Injected instructions. A fetched competitor page or a PDF in the input folder tells the agent to ignore its rules. Treat everything fetched as data, never as orders.
- Context bloat. Runs get slower and vaguer as the folder fills with finished work. Archive completed runs on a schedule.
- Silent tool failure. The agent reports success on a step whose tool returned nothing. Ask for the evidence file, not the claim.
Check dates, changes, sources, file size and evidence of completion. Inspect the actual output before approving the next step.
Which model should run which job?
Try different model options on the same task. A less expensive model may handle routine sorting, while a complex analysis may justify trying a more capable option. Include review and correction time when comparing cost.
| Model | Released | Per million tokens | Where it fits |
|---|---|---|---|
| Claude Fable 5.1 | Sept 1, 2026 | $10 in, $50 out | Long agent runs on cached context |
| Claude Opus 5.5 | Sept 22, 2026 | $4 in, $20 out | Large changes, audits and long coding jobs |
| Claude Haiku 5.5 | Oct 7, 2026 | From $0.10 in, $0.50 out | High-volume sorting and support drafts |
| GPT-6 Sol | Sept 22, 2026 | $2 in, $10 out | Hard research, long workflows |
| GPT-6 Luna | Sept 22, 2026 | $0.10 in, $0.50 out | Bulk drafting and classification |
| Grok 4.7 | Sept 21, 2026 | $2 in, $6 out | Long multi-step office work |
Reusing eligible input through caching can lower some usage charges. Anthropic lists Fable 5.1 cache reads at $0.25 per million tokens. The total saving depends on how much information is reused and what the rest of the task costs. Keep the software connections and their permission settings under review as well.
Do I need a developer to run agent workflows in a small business?
Some tasks can be set up without a developer. More complex connections or permission rules may need technical help. Someone in the business must own the process and check the results.
What should the first thirty days look like?
Choose one task and record results, mistakes and review time over several runs. Include incomplete inputs and exceptions before expanding or scheduling the process.
What does it cost to start?
The budget depends on the application, model usage, connected tools and review time. Measure a complete trial and check current provider terms before committing to ongoing costs.
Start with a workflow you can measure
Choose a recurring task with clear inputs and a person who can review the result. We can help define the first test, connect the relevant tools, and document the boundaries before expanding the workflow.