Which workflows should an agent run first?
Start where the work repeats, the inputs already belong to you, and a mistake shows up in a file before it shows up in public. That rules out invoicing and posting. It rules in the pile of preparation nobody enjoys: pulling research together, drafting page copy, checking a release, turning a process that lives in someone's head into a written one.
- 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
Anything that fails the last test belongs behind a gate or belongs to a person. Order matters more than tooling here. The businesses that wire up the sending step first spend the following month unwinding it.
The five workflows worth wiring this month
Each row is one workflow, the artifact it hands back, and the step a person keeps. The gate column is not a suggestion. It is the line we hold on client work, and it has not moved as the models have gotten better.
| 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 |
Notice what those gates have in common. Each one sits on an action that leaves the building: spend, publish, send, deploy, or a write into a system of record. Everything upstream of that line is drafting, and drafting is where an agent earns its keep.
What does one run look like end to end?
The wiring is deliberately boring. One folder holds the written job description, one holds the inputs, one holds the only place the run is allowed to write. The agent never writes into the input folder, so a bad run costs you a delete rather than a restore from backup.
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
The failures are predictable and they repeat across every business we have set this up for. Learn the shape of the five below and you can check for them in order, in about the time it takes to read the output.
- 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.
Run those checks in that order: dates, diff, sources, folder size, evidence. Four of the five are visible in the output folder before anyone outside the business sees the work, which is the entire point of gating on the artifact instead of the intent.
Which model should run which job?
Small businesses overpay by routing every step to the strongest model available. Match the model to the job and the bill stops being interesting. Volume drafting and classification go cheap; the judgment call and the final pass go to the strong model.
| 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 |
| GPT-5.6 Luna | July 9, 2026 | $1 in, $6 out | Bulk drafting and classification |
| GPT-5.6 Sol | July 9, 2026 | $5 in, $30 out | Hard research, long workflows |
| Grok 4.6 | Aug 12, 2026 | Unconfirmed | Long multi-step agent work |
Repeat runs are where cached context pays: Anthropic dropped Fable cache reads to $0.25 per million tokens with the 5.1 release, so a weekly workflow that re-reads the same brief costs a fraction of its first pass. Whatever host you use, the connection layer moved too. The MCP specification revision dated July 28, 2026 made the protocol stateless by default and tightened authorization, so any tool integration you set up before that is worth a look.
Do I need a developer to run agent workflows in a small business?
No, but you need someone who will actually read the output. The setup is a folder, a written job description, and one command. The real work is deciding which steps a person keeps, and that is an owner's call rather than a developer's.
What should the first thirty days look like?
Pick one workflow, run it weekly, and log what the agent got wrong each time. After four runs you will know whether the instructions need sharpening or the gate needs tightening. Add a second workflow only once the first one has run clean twice in a row.
What does it cost to start?
Agent Ready is free and covers the setup habits: folder structure, gates, and a written job description the agent can follow. A specialist skill is $149, and the business operations bundle is $447 against $596 bought separately. Model usage is billed by your provider on top of that.
Start with the workflow that already has a budget
The advertising and social skill runs the research, the variant matrix, and the platform spec checks, then stops at spend and publish.
Monolith AI Digital Marketing, Advertising & Social Media Agent Skill / $149, beta