Agents that can act. Within clear boundaries.

The concrete outcome

Agreed outcome: agentic workflows with bounded autonomy. lol.marketing delivers task boundary, tool rights, approvals, evaluation and shutdown in a form that can be reviewed, continued and operated internally.

Acceptance is not based on presentation impact but on agreed cases, quality criteria and control paths.

How the engagement works

• Record baseline, goal and owners
• Review data, permissions and edge cases
• Build and test with realistic cases
• Accept, document and hand over into operations

Price and boundaries

The lol.marketing fee is agreed individually after scope confirmation. Provider subscriptions, API charges, third parties, additional legal review and work outside the agreed scope are excluded.

Review state management for hosted agents

Hosted runtimes may store sessions, files and intermediate results. Anthropic Managed Agents is currently in beta and its documentation excludes Zero Data Retention and HIPAA Business Associate Agreement coverage. Sessions and uploaded files can be deleted separately through the API. A pilot therefore defines data needs, deletion, output verification and accountable owners.

Handle multi-step work with a defined objective

Some marketing tasks cannot be described completely as a fixed sequence. Research may need to inspect different sources, identify missing information and combine findings for a particular decision. An AI agent can help when the goal and permitted actions are sufficiently clear.

We develop agents for bounded marketing and sales tasks, including research briefs, analysis preparation and structured handling of variable information. We first establish whether an agent is necessary or a simpler workflow can already cover the job adequately.

Where the approach fits

The service suits teams that repeatedly handle open-ended work with an output that can still be assessed. A useful assignment names the audience, required information, permitted sources and intended use of the result. “Optimise our entire marketing” is too broad to provide that foundation.

An agent also needs a responsible person who can judge subject-matter quality. Where knowledge is missing or results cannot be checked, the task should first be narrowed. Additional independence makes sense only when its value justifies the extra operating effort.

What we develop

We define the task boundary, suitable information sources and allowed tools. The design includes limits on scope and duration, plus conditions for questions, approval and stopping. Outputs are arranged so someone can follow the underlying claims and results.

The agreed project includes working examples, technical implementation, representative checks and an understandable handover. We document which actions the agent may perform and which require a separate decision. Additional tasks or connected systems are separately defined extensions.

From idea to a tested agent

We start with a concrete assignment and suitable examples. Next, we examine information access and tools. A bounded prototype shows whether the approach produces useful results. Agreed improvements, further checks and introduction follow from that evidence.

An explicitly hypothetical assignment is to prepare a research brief for a new B2B offer. The agent reviews approved public sources, groups customer questions and marks unsupported assumptions. Its output includes traceable sources and unresolved points. Publishing campaigns, making contractual commitments or changing budgets does not follow automatically from that research task.

Requirements and pricing factors

We need defined tasks, permitted sources, appropriate access and a subject-matter owner. Confidential information and external actions are handled within the arrangements agreed for the particular use. Agent permissions are selected for its actual assignment.

Effort depends on task variability, tool count, information quality, evaluation requirements and support scope. Model and platform use can create ongoing charges. The proposal defines project scope and associated external costs. Error-free handling of every imaginable task is not a realistic service promise.

How does an agent differ from automation?

An agent can select its next steps more flexibly within an assignment. A workflow tends to follow known handovers. For predictable work, AI automation is often the more appropriate starting point.

Can an agent publish independently?

Such actions need explicitly agreed permissions and suitable checks. We often start with reviewable outputs or drafts. Publication is a separate decision within the particular process.

Will it work with our current tools?

That depends on available interfaces, permissions and required functions. We check feasibility beforehand. Tool integrations provide connections to existing systems where needed.