Updated September 28, 2026
10 min read
Tactics

Build an Agent Workflow: Research, Brief, Draft, Approve

Four Stages, One Artifact Between Each, and a Human Decision Where It Counts

Short answer

An AI marketing agent workflow works best in four stages: research, brief, draft, approve. The agent does the collecting and first drafts, a person signs off the brief before drafting starts, and nothing ships without approval. Each stage hands over a written artifact, so errors are caught where they are cheap.

GB
Written byGabor Barta— Co-founder, Oppira

Gabor leads product and content at Oppira. He has spent over a decade building tools and writing about competitive intelligence, social media analytics, and growth marketing for B2B SaaS companies.

Published September 28, 2026

1. What does a four-stage agent workflow look like?

Research produces a dated evidence note, the brief turns it into one decision, drafting turns the brief into channel copy, and approval is a person checking the draft against the brief. Two human checkpoints: after the brief and before publishing.

The four stages, what each consumes and produces, and who decides
StageInputOutput artifactWho decides
ResearchA bounded question and named sourcesEvidence note: findings with dates and linksAgent collects, person skims
BriefThe evidence note plus your playbookOne-page brief: goal, audience, message, proof, call to actionPerson approves or edits
DraftThe approved brief plus brand voice rulesChannel drafts with the brief line each one servesAgent drafts, person picks
ApproveDrafts plus the briefApproved post, scheduled or publishedPerson, always
The four stages, what each consumes and produces, and who decidesThe artifacts are the point. When something goes wrong, you can see which stage introduced the error instead of rerunning the whole chain.

The shape matters more than the tools. You can run this with a chat assistant and a shared folder, with an agent in a terminal, or inside a marketing platform. What makes it work is that no stage starts from a vague instruction, and no draft is written before a person has agreed on what it is for.

2. How should the research stage be scoped?

Give the agent a closed question, a named list of sources and a time window, and ask for findings with dates and links. Open questions like find everything about our market produce confident summaries of whatever was easiest to reach.

A good research request reads like this: what did Veltrix and Sondera publish about pricing between September 1 and September 28, from their websites, their LinkedIn company pages and their current Meta ads, with one line per finding and a link. It has names, sources and a window, so the agent can finish and you can check the work.

The failure mode is well documented: an agent cannot tell you what it did not find, so an open question returns a partial picture presented as a complete one. The guide on delegating research to agents covers where the line sits in more depth.

Ask the agent to structure the evidence note as:

  • Finding, in one sentence.
  • Source link and the date it was published or observed.
  • Whether it is a direct quote, a number or the agent's interpretation.
  • A short list of sources it tried and could not reach.

3. What goes into the brief, and why does a person approve it?

A goal, an audience, the one message pillar the content serves, the proof that supports it, and a call to action. A person approves it because the brief is where strategy gets decided, and deciding strategy is not the agent's job.

A fill-in brief template the agent drafts and a person approves
FieldWhat to writeExample
TriggerThe finding from research that prompted thisVeltrix now leads its Meta ads with its entry plan cut from $39 to $33 a month
GoalWhat the content should change, in one lineKeep switching intent from turning into price comparison
AudienceWhich ICP, at which stageOperations leads evaluating tools this quarter
Message pillarOne pillar from your message house, by nameSetup in a day, not a month
ProofEvidence you are allowed to use, with its sourceMedian onboarding time from your own product data
Call to actionThe single next stepBook a fifteen-minute setup walkthrough
Channels and countWhere and how many piecesTwo LinkedIn posts, one email, one ad variant
Not sayingClaims to avoidNo direct price comparison, no naming Veltrix
A fill-in brief template the agent drafts and a person approves

The approval is quick, often two minutes, but it is the most valuable two minutes in the workflow. Approving a brief is cheap; discovering at the draft stage that the angle was wrong costs every draft. The Not saying row is where most brand and legal problems are prevented.

4. How do I keep drafts on brief?

Give the drafting step only the approved brief and your voice rules, ask for two or three variants per piece, and require each draft to name the brief line it serves. Drafts that cannot point to the brief are off brief.

Narrow input is the lever. A drafting step that sees the whole research note will drift toward whatever was most interesting in it. A drafting step that sees only the approved brief, the brand voice rules and your content policy lines will write to the pillar you chose.

Instructions that keep drafts honest:

  • Write two or three variants per piece, each with a different hook, all serving the same pillar.
  • Start each draft with the pillar and proof line it uses, so the reviewer can check it at a glance.
  • Use only the proof listed in the brief. If a claim needs a number that is not there, write [NEEDS SOURCE].
  • Respect the channel: length, format and any character limits for the platform.

5. What does the reviewer check before approving?

Five things: it serves the brief, every claim is backed by the listed proof, it sounds like you, it respects the Not saying row, and the link and timing are right. If any fails, send it back with the specific line.

  1. Check it against the brief. Read the pillar line at the top of the draft, then the draft. If the draft argues something else, reject it rather than editing it into shape.
  2. Check every claim. Each number, customer outcome or comparison should trace to the proof field. Anything that does not is removed, however good it sounds.
  3. Read it aloud for voice. Fluent is not the same as on-brand. Listen for phrases your team would never say, and for the generic openers AI drafts default to.
  4. Check the Not saying row. Competitor names, price comparisons and restricted claims are where approvals matter most. This is the step that protects the brand.
  5. Check link, tracking and timing. Confirm the destination, the UTM tags and the scheduled time, then approve. Publishing is the only step that cannot be undone.

6. Where do agent workflows usually break?

At the handoffs: research that answers a different question, briefs approved without reading, drafts that see too much context, and approvals that become rubber stamps. Each has a simple fix.

Common failure points and the fix for each
WhereWhat goes wrongFix
ResearchAnswers an easier question than the one askedName sources and a window, ask for sources it could not reach
Research to briefInterpretation presented as factLabel each finding as quote, number or interpretation
BriefApproved without readingKeep it to one page and require an edit or an explicit yes
DraftDrifts to the most interesting research detailFeed only the approved brief and voice rules
ApproveBecomes a rubber stamp after a few good weeksKeep the five checks, rotate the reviewer occasionally
Common failure points and the fix for each

Oppira runs this loop inside one system: Intel supplies the dated research, the agent proposes playbook edits and content plans, Studio drafts in your brand voice to your message pillars, and nothing is published or changed without your approval. The stages are the same whichever tools you use.

7. How do I start without rebuilding everything?

Pick one recurring content type, run the four stages by hand for two weeks with the templates above, then automate the stage that took longest. Usually that is research.

Choose something you produce every week, such as a LinkedIn post that responds to market news. Run the stages manually with an assistant, saving each artifact in one folder. After two weeks you will know which stage takes longest and which produces the most rework.

Automate in that order. Research is usually the first candidate because it is collection, which agents do well. Briefing is the last candidate, because it is where your judgment lives, and a workflow that automates the judgment away is just a faster way to publish the wrong thing.

Key Takeaways

Four stages, four artifacts

Evidence note, brief, drafts, approved post. Written handoffs make errors visible at the stage that caused them.

Closed research questions

Names, sources and a window let the agent finish and let you check. Open questions return partial pictures presented as complete.

A person approves the brief

The brief is where strategy is decided. Two minutes there saves rewriting every draft.

Drafts see only the brief

Narrow input keeps drafts on the chosen pillar instead of the most interesting research detail.

Approval has five checks

Brief fit, claims, voice, the Not saying row, and link and timing. Publishing is the only irreversible step.

Automate research first, briefs last

Collection suits agents. Judgment is the part worth keeping.

Frequently Asked Questions

Oppira

Turn reading into a reaction

Oppira watches your competitors, keeps your playbook current, and drafts the response. See your market clearly by tomorrow.