REFERENCE PATTERN / ILLUSTRATIVE
- Research
- Knowledge / draft
- Programmatic checks
- Human approval
- Publish
A multi-agent system is not automatically a useful business system. Giving several language models different roles may generate activity, but operational value comes from how work is specified, checked, handed off and recovered when a step fails.
The Jiewen GEO Optimization System combines a Python/Flask foundation, structured knowledge, AI-assisted research and drafting, review processes and Playwright-based publishing tasks. The architecture illustrates a broader lesson for business automation: agents need explicit boundaries, and deterministic software must remain responsible for the things that can be checked precisely.
This article describes engineering patterns informed by that project. Example schemas and workflows are illustrative, not exports of proprietary source code or guarantees of unattended publishing.
Start with a workflow, not a cast of agents
A team might need to turn source materials into a reviewed article and publish it through a permitted channel. That goal already suggests distinct stages: collect source material, structure evidence, draft, verify, approve and publish.
Only then should we ask which stages benefit from a model's judgment. Research synthesis and drafting may. Checking that a URL is allowed, required metadata exists, a file is valid or a human approved release is often better handled by deterministic code.
Calling every step an "agent" obscures accountability. A useful implementation makes each task's input, output, owner and failure behavior visible.
The shape of an auditable pipeline
Think of the system as a series of contracts rather than a conversation that never ends.
Research: record provenance
The research stage should capture source locations, timestamps where relevant, and what evidence supports each claim. It should not silently turn a model's inference into a verified fact. Source retrieval failures must remain visible to downstream stages.
Knowledge: normalize what is known
Structured context helps prevent contradictions between drafts. A company name, product category or approved capability can be represented as a controlled field rather than repeatedly recreated from an open-ended prompt.
Draft: produce a candidate, not a release
Generation can transform verified context into prose. It should also label uncertain claims for review. A good interface makes the draft easy to revise instead of treating the first model output as final.
Quality gate: make failures actionable
Programmatic checks can validate schema, links, unsupported dates, missing citations, format rules and forbidden claims. Model-based review can help spot softer problems, but it should not replace clear validators when a requirement is machine-checkable.
Approval and publishing: control side effects
Publishing changes external systems. A browser automation task can fail halfway through a form, lose authentication or encounter a layout change. Retries need to be idempotent where possible and must not create duplicate public posts. A human approval checkpoint and an audit trail are often more valuable than an unsupported promise of universal one-click posting.
Why task state matters
A business workflow should know whether a job is queued, running, waiting for review, failed or complete. Without explicit task state, it becomes difficult to answer simple operational questions: Which output was approved? Which input version was used? Did the publishing step actually succeed? Can we retry safely?
A compact illustrative record might look like this:
{
"job_id": "example-001",
"stage": "review",
"source_revision": "brief-v3",
"draft_revision": "draft-v2",
"checks": {"links": "passed", "claims": "needs-review"},
"approved_for_publish": false
}
The key point is not this exact JSON structure. It is that the workflow can stop, explain why, and resume without pretending the work is finished.
Where Playwright fits — and where it doesn't
Browser automation can bridge services that lack an appropriate API, but it inherits the fragility of web interfaces. Selectors change, sessions expire, consent dialogs appear and platform policies differ. A Playwright-driven publishing module therefore benefits from explicit preconditions, narrow permissions, bounded retries and post-action verification.
In the Jiewen case, the publicly described system uses Playwright within a larger Python workflow. That does not imply every platform is supported, or that all publishing can be performed without human review. A trustworthy product states those boundaries plainly.
Multi-model support without multi-model confusion
Jiewen's case notes refer to integrations with DeepSeek and Qwen. When multiple models are available, a workflow can route tasks according to the requirements of the stage rather than assume one model is optimal for everything.
However, changing the model is not the same as changing the truth standard. The system should keep source material, data validation and human approval independent from any individual provider. Model outputs are candidates for controlled processing, not authoritative statements about business performance.
Security and operational guardrails
Never put unrestricted production credentials in an agent prompt. Use least-privilege integrations and isolate publishing permissions from research permissions. Keep sensitive inputs out of logs by default, set execution limits and record enough sanitized diagnostics to investigate failures.
For content-oriented GEO workflows, another boundary matters: optimization is not a guarantee of placement in search engines or AI-generated answers. Results depend on independent platforms. A credible engineering case should describe implemented functionality rather than promise rankings, impressions or leads that have not been measured.
The important design decision
The useful split is between probabilistic work and deterministic control. Let models help interpret and draft; let typed state, validation, authorization and explicit approvals govern side effects. Multi-agent orchestration becomes a practical system when it can explain what happened, recover safely and prove what was published.
Related engineering work: Jiewen GEO Optimization System case study. For a workflow that connects models, APIs and real operational steps, start a conversation.