Chat interfaces pollute their context, make precise editing difficult, and rarely force users to clarify intent.
2
Documents give people a structured way to describe complex systems and create repeatable agent assignments.
3
Background agents need human approval, explicit authority, communication protocols, and documented workflows before they can handle important work.
Summary
Filip Kozera argues that chat is a poor interface for building reliable systems. Long conversations accumulate irrelevant context, make version control difficult, and do not force users to answer the questions that would clarify their intent. Documents work better because they let people group ideas, specify details, revise individual sections, and explain a system to someone who does not already know how to build it. Kozera connects this document-based approach to background agents. A conversation can help brainstorm, but a repeatable process should be written down before it is connected to a trigger such as a cron job, Gmail event, or implicit event. The agent can then work in the background and ask for approval when it struggles. Kozera expects people to manage groups of agents, with protocols that describe data access, authority, constraints, and human approval. He is also direct about the risks: companies will adopt agents slowly when mistakes could have serious effects, and humans will remain responsible for the agents acting in their name.
Chat interfaces create polluted context and make precise iteration difficult
Kozera says long conversations with Claude or OpenAI often become workflows built from projects and copied text. As the context window fills, useful instructions sit alongside what he calls "gibberish and garbage." This makes it hard to change one sentence precisely, maintain version control, or reuse part of the work. He also lists model laziness as context grows, the lack of logical grouping and nesting, and the fact that chat exposes only one abstraction layer. Users cannot easily choose between specifying every detail and setting up a higher-level process.
Documents force people to explain complex systems clearly
Kozera presents documents as the long-standing way humans specify systems for people who do not already know how to build them. He gives a deliberately uncertain example of a product requirements document dating back to Noah's ark, "around three and a half thousand years ago." A document gives users a place to organize related material, revise exact wording, and describe a system at different levels of detail. That structure creates the clarity he says chat interfaces often fail to demand.
Background agents extend handcrafted workflows toward general decisions
Kozera describes a progression from manually built workflows, such as Zapier-style automation, to specialized agents that make one or two fuzzy decisions. His diagram places workflows with high importance and high occurrence in the handcrafted category. General agents are only beginning to take decisions, and people are reluctant to give them control when the work matters a great deal. Background operation is the part of systems such as Manus and Deep Research that points toward this next model.
Human approval gives agents a controlled path to useful work
The proposed remedy is a human in the loop. An agent can do a large amount of work, while a person approves or rejects the result, changes how it was produced, or fixes the agent's logic. Agents can start from explicit triggers such as a sent email or Slack message, or from an implicit event such as a meeting with an investor that should update a CRM. Once many agents operate together, people need protocols for controlling outputs and for letting agents exchange information.
Agent systems will need explicit authority and communication rules
Kozera expects background agents to move from consumer tasks such as managing email into organizations with internal and external agents. One agent might create Jira tickets for engineers, while another consults a general agent about information in Notion. He says the Model Context Protocol is an early protocol, but it does not provide enough information about an agent's constraints, authority, or need for human approval. Companies will adopt these systems slowly when they need verified tools, correct permissions, and confidence that agents will not cause serious damage.
People will manage agents by setting intent and reviewing their actions
Kozera says humans will need to put their own taste and intent into agents and take responsibility for what those agents do. He compares the current situation in coding to managing a team of interns. Engineers who can both do individual work and manage a group of AI agents can gain more from the tools. Some strong individual contributors reject AI because its code does not meet their standards. For Kozera, the ability to set a quality bar and direct a swarm is part of the work.
A document should define a repeatable process before an agent runs unattended
Kozera compares chat to brainstorming. After brainstorming, people still need to write the document that explains what an agent should do. This is appropriate for repeatable processes that have earned trust. The written assignment can then connect to a cron job, Gmail trigger, or implicit trigger. The agent works in the background, with latency mattering less, and surfaces problems only when it struggles or needs approval. Human work shifts toward creating assignments, expressing taste, and reviewing results.
Agent tools may become the lasting advantage for enterprises
Kozera expects consumers to adopt background agents first, followed more slowly by enterprises. He predicts that companies will build specialized agentic tools using the newest models, while the tools themselves become the source of durable advantage. His proposed workflow leaves people reviewing and editing outputs as agents take on more background tasks. He presents this as a prediction rather than a settled result, ending with, "Let's see if it's true."
"I realized that a lot of these things in the context window are just like gibberish and garbage and so we get context pollution."00:24
Who should watch
You are building agent workflows in chat and need a clearer way to turn experiments into repeatable processes.
Your team wants background agents, but needs a practical model for triggers, human approval, permissions, and review.
You manage engineers or AI tools and want to understand why setting a quality bar may matter as much as writing code.
