About ten years ago I devised a framework. Four words. Context, Clarity, Closure, Consistency. I was trying to solve something practical: why do smart people, with real resources and good intentions, still produce outcomes that feel broken? The answer, I kept finding, was never the idea. It was the infrastructure around the idea. The connective tissue that holds the thinking together when the thinking gets hard.
A decade later, I’m watching the industry relive the pain through Agentic Harnesses. Context, Memory, Evaluation, Reinforcement. A structure for holding intelligent behavior together when the intelligence gets hard. The parallel is so structurally similar that I keep thinking: we weren’t designing a human performance framework back then. We were, without knowing it, describing the minimum viable architecture for any reasoning system that has to produce reliable outputs in an uncertain world.
Context: The word that could mean everything, and nothing at the same time
Context means the shared understanding of why we are doing what we are doing. And not just the brief, but the full picture: the history that shaped the problem, the constraints in people’s heads, the organizational gravity. Most project failures happen in the first twenty minutes. Not because people make bad decisions. Because people make decisions inside different mental models of the same situation, and nobody ever noticed. I have tried to highlight this difference of contexts with Part 2 of my Build vs Buy argument – the Trifecta of…
In an agentic harness, Context is the information window available to the agent at the moment of Action. The prompt, the MCPs, the retrieved knowledge, the tools, the infrastructure. It is whether the agent has enough of the picture to act correctly.
A human team with bad context makes decisions that are locally reasonable and globally wrong. An agent with bad context does the same thing, faster, at scale, with more confidence. And the thing that makes context hard in both cases is not access to information. It is selection.
The problem was never quantity. It was always relevance.
British Hungarian philosopher Michael Polanyi summarized the context problem by the statement “we can know more than we can tell”, famously known as the Polanyi’s paradox.
A 2011 study(1) about tacit and explicit knowledge uncovered the differences in the light of medical professionals.
In 1990 (2) a Stanford University study by Elizabeth Newton, a group of subjects were asked to tap out the beats to a song on a table with just one finger, while another group was asked to identify the melody. The tappers would always overestimate how many of the songs would be identified by the listeners. The tappers are so familiar with the song that they always assumed listeners would know.
In an agentic harness, the equivalent is retrieval architecture. RAG pipelines, vector search, context compression. Different tools. Same design intent.
Clarity translated into Action
Clarity is precision of intent. Clarity makes the difference between a goal that sounds clear and a goal that actually is. Clarity could mean the brief that everyone nods at, versus one that everyone could execute independently and arrive at the same place.
I’ve seen smart teams fail at building Clarity in the same way. They would align on the what and leave the how-well vague. Most often, on purpose. Specificity is a form of commitment, and commitment is a form of risk that you have not yet seen. So, they chose to stay fuzzy. And fuzziness felt like flexibility until everyone built something slightly different and nobody was technically wrong.
The agentic harness has Actions. Actions define what each system is optimized for.
Clarity is front-end: what are we trying to do, and do we all agree on what done looks like?
The human solution is discipline. The machine solution is architecture.
Evaluation to Closure: Knowing when you’re really done
Closure is the practice of deliberately ending things. Projects, decisions, debates, phases. Bringing them to an explicit stopping point, extracting the learning, and releasing the energy tied up in the open loop. Closure is uncomfortable. Most organizations are excellent at starting, reasonable at doing, and catastrophic at finishing. Not finishing as in shipping the thing. The product ships, but the post-mortem doesn’t happen. The handover doesn’t happen. The loop stays open.
In the Agentic Harness, the equivalent is Evaluation: the feedback mechanism by which a system determines whether what it just did was good. Evaluation can be automated, human-in-the-loop, or self-assessed. Without it, the agent repeats its errors with perfect consistency because it has no signal that they were errors. One of my pet peeves – observability lands in this zone.
Both are systemically underinvested. Closure gets cut when deadlines compress because the next thing always feels more urgent than reflection on the last. Evaluation gets simplified when compute costs matter, so teams use proxies, or evaluate at the end instead of throughout, by which point the errors have already compounded.
Reinforcement to Consistency: Gearing up for the long game
The fourth component is where both frameworks reveal their deepest ambition. This is not what the system does in a single interaction, but what the system becomes over time.
Consistency is the hardest as a human. It means the principles and standards the team holds do not change based on convenience. The moment consistency gets expensive; exceptions are made. Each exception feels locally justified. Cumulatively, the exceptions become the rule, the rule becomes decoration, and people stop trusting the framework.
Reinforcement happens through memory. Whether from human feedback, AI feedback, or environmental reward signals, it is the system’s answer to: what gets rewarded here? What patterns does the agent strengthen because they produced good results and which does it abandon because they did not?
Memory is the mechanism for persistence across interactions. Short-term memory lives in the context window. Long-term memory lives in external stores: databases, vector indices, episodic logs.
You can have perfect Context, Clarity, and Closure in a human system and still produce an inconsistent organization if the reward signals are wrong. And you can have perfect context retrieval, memory, and evaluation in an agentic system and still train a bad agent if the reinforcement signal is unspecified.
The Structural Isomorphism
4C was built for humans inside organizations. The Agentic Harness was built for AI systems inside software pipelines. The substrate is different, but the parallels are there.
Both start with the same premise: intelligent behavior does not emerge from intelligence alone. It emerges from intelligence plus structure. A brilliant person with no context, no clarity about goals, no feedback loops, and no consistent standards produces brilliant but unreliable outputs. A powerful language model with no context architecture, no memory, no evaluation, and no reinforcement produces the same.
Context, Clarity, Closure, Consistency. Context, Action, Evaluation, Reinforcement. The names are different. The structure is the same.
The machine however has an edge that 4C for humans only could dream of as the north star state.
Context, not as concept but as implementation. In the human world, organizational context is one of the most persistently broken things. The knowledge lives in people. People leave. You rebuild it. They leave again. The organization cycles through the same hard lessons every three to five years because institutional memory has no real architecture. It lives in email threads and shared drives and the heads of people who were there, degrading at roughly the same rate as their tenure.
Evaluation at scale. In the human world, the feedback loop is slow. A retrospective happens weeks after a project ends, if at all, and the learning diffuses imperfectly. In an agentic system, evaluation can happen at every step, across millions of interactions, aggregated faster than any human review process. The learning surface area is orders of magnitude larger.
The architecture can hold. The question is whether the people building it will hold to the architecture with the same discipline that the architecture demands.
The people who have spent years building human performance frameworks are not behind.
The engineers building agentic systems are largely solving infrastructure problems. How do you build a reliable memory store? How do you design a reward model that generalizes? The first-order problem, however, is: what does the system need to do, for whom, in what conditions, with what definition of success? That is a design problem, a human performance problem.
The convergence is already starting. The best agentic system designers are the ones who have internalized both traditions. They speak both languages. And they are rarer than anyone in either camp is willing to admit.
Ten years ago, when I imagined 4C for people, I thought I was solving a human problem. Turns out it was a reasoning problem.
We are building systems that reason at scale, act autonomously, and produce outputs that shape real decisions in real organizations. The question of how to make those systems reliable is not a technical question alone. It is a design question. The people who built frameworks for the first substrate have something to teach the people building for the second. And the people building for the second have solutions to problems the first substrate could never fully solve.
That conversation is the most interesting one in technology right now.
I have been having a version of it for ten years. It is good to finally have more people in the room.
References
(1) Kothari, A.R., Bickford, J.J., Edwards, N. et al. Uncovering Tacit Knowledge: A Pilot Study to Broaden the Concept of Knowledge in Knowledge Translation. BMC Health Serv Res 11, 198 (2011) Uncovering Tacit Knowledge: A Pilot Study to Broaden the Concept of Knowledge in Knowledge Translation – PMC
(2) Newton, Elizabeth Louise, Ph.D.: Rocky Road from Actions to Intentions
Leave a Reply