The Two Extremes
AI is already changing the way how we work. The expectation was that it would offload human work. In reality, many teams just ended up with getting more work. Many people still feel that the only tasks they can reliably trust AI today are being the question answering bots, who help you search and synthesize, and doing writing and coding with you. These question answering bots still expect you to know what to ask, stitch everything together manually, and push work to the finish line. The promised business value that AI will bring seems still a long way to go.
On the other extreme, many AI companies are selling the beautiful dream of fully autonomous AI agents can complete complex tasks for you without human participation (here we are not talking about non-cognitive simple tasks). It may come true for some tasks eventually, but it may never come true for many tasks. I don’t think you are really comfortable to let the agents being autopilot to make big business decisions.
Neither the 2 extremes are a good answer to really offload human and bring business value. The answer to that is obvious, the middle ground. AI agents that are proactive and embedded in your environment, that offloads real work, but stays below full autonomy so humans still make the calls but only make minimum effort.
Context is the key
Context for AI Agents
It’s now a consensus that context is the key to making AI agents actually useful. A generic model with no context is just a clever chatbot. An agent that can see the relevant part of your data, systems and history at right time can start to behave more like a superhuman worker inside your environment.
This is context engineering, which has become a hot topic in the industry. Context Engineering is the work of deciding what an agent should see before it responds, and how that information is delivered.
In practice, this means:
- Giving agents access to the right sources: databases, logs, dashboards, tickets, documents, CRM, calendars, code.
- Retrieving the most relevant pieces for the current task.
- Structuring that information into a compact “working memory” the model can use.
- Enforcing the right constraints (policies, limits, tools it is allowed to call).
If your agent system can find and feed the model with the most relevant information for the current task, recent events, current KPIs, historical performance, previous decisions, it stops behaving like a question answering machine and starts behaving like someone who has read all your internal information, checked all the systems, and then gives an analysis.
This is all about explicit context: everything we can write down, store in a database, expose via an API, and systematically feed into the agent. It’s necessary, but, as we’ll see next, it’s not sufficient.
Humans have implicit context
We can theoretically write down all the knowledge and information we have in our brain to documents and databases, where AI agents can access them, so that AI share the same context we human do. Isn’t that correct? Maybe not really.
We know more than we can tell. The information we can easily output to somewhere externally outside our brain we can call it explicit context. You can try to teach another person everything you know, but they still won’t have the exact same understanding. Some things are lost in communication, and some knowledge sits behind what you can easily explain. Many of those are instinct reactions, feelings and thoughts, they were synthesized from long history of life and work experience and knowledge. We can call it implicit context.
In a real business, you can imagine how you can record the following examples into a document or database:
- A sales lead knows that a prospect who asks a certain question in the first call is 99% not going to buy, even though the CRM history alone wouldn’t show that pattern.
- In calm times, the team watches conversion rate; in a crisis, everyone quietly switches focus to time-to-recovery and refund volume even though it’s not the “north star metric”.
- For a certain product service, you learned that 5% error rate is actually fine on one endpoint (because of a fragile external dependency), but 0.1% is unacceptable on another because any failure there leads to on-call at 3am.
- The wiki says “open an incident in tool X”, but in reality you ping a specific Slack channel first or nothing moves.
This kind of context lives in people’s relationships, scars, and long-term mental models. It’s the type of knowledge that’s hard to turn into a neat schema or embedding, and that’s why fully autonomous agents will hit a wall here.
Even you can gradually externalise some of these implicit knowledge, and either bake them into the model weights or save them to a database, but these information will most likely be buried in the sample distribution.
Humans remain the irreplaceable node
Because of the many crucial implicit context information, and because majority of the physical real world is still not fully digitalised, humans are still the irreplaceable node in many scenarios.
Wait, isn’t AI supposed to offload humans?
But how? Let human and AI collaborate, but not in the way that human being the stitchers. We should let humans do what can only be humans to do, and let AI do what they are really good at.
Human should only participate in the place and timing that they really need to, where the crucial decisions should be made, where only they can fill in the context gap, where they are the keys that unlock the AI agents when they reached their context limit.
What Humans should not be doing
Given the current capabilities and foreseeable improvements of the models, a lot of tasks human should not be doing. There are obvious tasks that AI is way much better than human like information retrieval, extraction and synthesize, SQL queries, report writing, etc.
There are also tasks that are not very obvious, but humans should not be doing, examples are:
- Human should not keep gathering and giving context information from different systems to AI, for which the AI agents should be able to get explicitly themselves.
- Human should not be acting like scheduled jobs to check dashboards every morning, compare KPIs, scan for anomalies, watch for time-based events.
- Human should not spend more time on navigating and interpreting information than actually deciding what to do.
- Human should not keep entering various systems to makes changes after decisions are made.
These tasks are not unique or valuable that only humans can do reliably with implicit context. Human should spend more time on thinking, prioritizing and deciding.
What AI Agents should be doing
So what should the “middle ground” look like? What should AI agents do to offload human and let human focus on the tasks only they can do?
In short, an agent system has access to the same data and systems as human do, it should be:
- Watching
- Navigating
- Aggregating
Not deciding- Acting
Watching
AI agents should continuously monitor data updates, metrics, logs, events, etc. They should not wait for a human to come and ask, “Is everything okay” or “What has happened?”
Navigating
Grant AI agents access to different systems (read-only for now). Let them navigate across tables in your databases, documents in your information systems, and other tools to build a coherent picture of what’s going on, instead of making a human click through five dashboards, write tens of SQL queries, read through documents, search big news.
Aggregating
This will be much more powerful than conditional triggers that you can set on your data metrics. The agents can oversee signals from various sources to make comprehensive assessments on if everything is okay, instead of only comparing a few KPIs to their thresholds.
Not Deciding, but suggesting
As aforementioned, AI may miss important implicit context, and of course it may also miss information that it should have navigated to.
For anything with real business impact, it should pause and ask for a human decision.
Additionally, AI agents are not entities yet, they cannot take responsibilities. Only the ones that will take responsibilities should pull the triggers. However, AI agents can support human to make way much better decisions.
Given comprehensive explicit context and totally logical reasoning, AI agents can provide human 1-5 decision suggestions, with support evidence, possible outcome, etc.
Acting
Once the decisions are made by humans, the agents should go execute directly . From simple cases like updating a value in an internal system (e.g., price adjustment), to update Google and Meta Ads campaign with adjusted budget and CTA, to create a customer survey plan, and to update the code base to do A/B experiment for the visual design on improved customer experience.
This flow is almost the definition of human-in-the-loop. For anything with real business impact, it should pause and ask for a human decision. Over time, for some low risk, reversible, or high volume tasks, you can gradually let the agent execute directly based on past human consent, moving from human-in-the-loop towards human-on-the-loop.
From coding agent to business agent
Coding agent is the most advanced example of human-agent collaboration, there are various reasons
- Result is relatively easy to validate and evaluate,
- High volume data of code and tests to train model,
- The context of the code is mostly contained within the project repository,
- Developers are the fastest ones for adopting new tools and new ways of working, so it evolves faster with usage feedback.
In practice, there are various work flows that suit for different type of coding tasks, where developers and coding agents work together in a structured way. One of the typical flows to implement relatively complex features looks like this:
- The developer describes the task in natural language (“add a feature flag for X”, “fix this bug in the checkout flow”),
- Lets the agent scan the repository to identify relevant files, search on the web for required information or existing implementations, and then propose a plan,
- The developer reviews and edits that plan, then asks the agent to implement it,
- The agent generates code changes, runs tests or linters, and shows a diff,
- The developer reviews the diff, adjusts where needed, maybe asks the agent to refine parts, and only then commits or opens a pull request.
This iterative “plan -> generate -> review -> refine -> approve” loop means the coding agent does most of the work, while the human keeps control over intent and final approval.
In business, things are relatively more complex/non-straightforward comparing with coding. However, if we design and implement a good agent system, agents in the business will catch up faster and provide values.
- Result is not easy to validate and evaluate, but human can feed the result and analysis back to the agent system for future improvement,
- The models are not fine tuned for every specific business scenario, but they are already pre-trained with a broad business concepts and reasoning patterns. And you can close much of the gap by grounding them in your domain with context engineering, and human feedback loops, so the system “learns” your definitions, KPIs, decision patterns, etc.
- The context is not well contained at one place, but human can grant access for the agent to navigate, and gradually structure them for agents to access easily,
- As human use agents more, there are more real-world inputs to iterate how the agent system should be designed and implemented for that specific business domain.
Closing
Back to the main theme of how AI agents should offload human. We don’t want AI that only replies to my prompts. We also don’t want AI that silently makes decisions we will be blamed later.
We want agents that:
- Notice problems or opportunities before we do,
- Do the legwork we’d rather not do,
- Line up sane options,
- Act for me once I made decisions.
That’s the design philosophy we use at Actorise.
Models will keep improving, even before a new model architecture or a new type of inteligence is discovered, the scaling law is still taking effect. The capabilities of AI agents will keep evolving.
But don’t be scared.
As human, our personal experience makes us unique. We carry a unique history context and unique ways of interpreting the world that no model has.
AI agents are not replacing human, they are forcing us to think on a higher dimension.