The six skills of agentic coding every developer needs to master

AI Coding: A New Coding Paradigm

When computers first arrived and changed the face of technology, the methods to code them involved directly speaking their language (binary) - flipping switches that corresponded to bits of 0 and 1. Soon, it became obvious that for any more complex logic, we needed a translation layer that converted a higher level language into binary - assembly. Likewise, it again became clear that many of these operations were repetitive and could be packaged into even more higher level statements, which became languages like Fortran and C. In the years since, even more abstracted languages like python and rust have taken over. This is what we can refer to as the "deterministic abstraction" stack - where the language is just packaging assembly commands that can be evaluated. With the advent of LLMs, we now have the capability to abstract out semantic, not just deterministic, understanding. This has given birth to a new kind of programming language so to speak - agentic coding agents - which take natural human language, the language of thought, and instructs a computer to write code to that specification. This has immense untapped potential, as now any idea at our fingertips (with the right coding agent) can be coded up within a day.

The Strengths and Weaknesses of AI Coding

As with any technology, using it in different ways and methods can produce different outcomes - for example, a welder must learn to point the flame at certain distances and points for maximum effectiveness. In the case of AI Coding, this comes back to the engine underlying the agent coding tools - the Large Language Model, or LLM, built off of the Transformer neural network model.

Transformer neural networks are a form of feed-forward neural network (machine learning) model, that fundamentally take in a fixed size input at each iteration, and output a collection of output words. Because of the nature of the transformer, it only has access to the words in its context window, as well as its training set, in which to make a decision - in this case, the decision is "what words do I output in response to the user". This means that for example, knowing what to input into the limited context window is fundamental to the best outputs from the models. This is the origin of one of the skills, which measures how well people choose what and what not to include in the message that gets sent to the user.

LLMs are probabilistic engines that don't quite reason the same way as humans - they output likelihoods of the next tokens given the prior ones in a sequential fashion. This means they are prone to what is known as "hallucinate", or create potentially random text or code output that wasn't asked for. This means it is incredibly important to verify that the output matches what was requested. This leads to another of our scores, which asks if the user performed baseline due diligence in checking what was outputted, before shipping the code in the repo.

What AI Skills Measures

In the same way that programming languages required skills and techniques to master, there is no reason to believe that agent coding also doesn't involve its own set of skills and techniques to master. Here at Weave, we have developed what we believe is the first "AI Skills" set, which we judge based on users conversations, that can help train and guide people towards becoming better and more productive agent coders.

AI Skills evaluates how AI-assisted engineering work unfolds, not only the code produced at the end. It treats the conversation as evidence: user prompts, agent responses, tool-use summaries, verification attempts, handoffs, and revisions made when new information changes the task.

The practical question is simple: did the agent work in a way that was planned, verified, context-aware, collaborative, parallel when useful, and still under human engineering ownership?

What is being Measured

Weave's AI Skills reviews a person's conversation with a coding agent and evaluates how the work was carried out across the six Skills below. The object being measured is not the person's ability in the abstract, and it is not simply whether the agent eventually produced working code. It is the observable practice of using an agent to do engineering work: how the person sets direction, supplies context, responds to evidence, coordinates work, and decides what to trust.

This distinction matters because the same final patch can be reached through very different conversations. One person might give the agent a clear goal, identify the important constraint, ask for the right checks, and redirect the work when a result does not fit. Another person might arrive at the same patch by accepting whatever the agent proposes without examining the assumptions along the way. Looking only at the final code would treat these conversations as equivalent. AI Skills treats the conversation as part of the engineering artifact, because it contains the decisions and evidence that produced the code.

Why these skills were chosen

These Skills were chosen through internal evaluation at Weave with engineers and researchers who use coding agents extensively, together with a review of the emerging practice of AI-assisted software development. We looked for the behaviors that repeatedly changed the quality of the work, regardless of the particular agent, model, language, or repository.

The result is a set of foundational practices rather than a checklist for one tool. A person does not need to use every behavior in every conversation. A small, clear task may not need multiple agents or a complicated plan. The question is whether the person recognized the moments when a behavior mattered and handled those moments well.

Together, the Skills describe the control loop of agentic engineering. Planning sets a direction, context gives the agent the information needed to act, Verification tests whether the work can be trusted. Collaboration and parallelization organize how people and agents work together. Ownership keeps the consequential decisions with the engineer.

The Six Skills

Planning

Planning measures whether the person turns an intention into a path the agent can execute. This includes stating the desired outcome, preserving important constraints, identifying decisions that cannot be left implicit, and breaking non-trivial work into useful steps. Strong planning is not the longest plan. It is a plan that exposes the decisions that matter and changes when new evidence invalidates an earlier assumption.

Verification

Verification measures whether the person makes the work answer to evidence. This includes asking for checks that match the risk, examining what those checks actually observed, and changing direction when the result is incomplete or wrong. A successful command is not automatically a successful verification. The relevant question is whether the conversation establishes what was checked, what it showed, and what remains uncertain.

AI Context Operations

AI context operations measure whether the person gives the agent the context it needs and keeps that context coherent as the work continues. This includes scoping an instruction, carrying forward a decision, reconciling a new source with an earlier belief, and filling an information gap before it becomes a wrong implementation. Context is an active part of the work: too little context causes drift, while unstructured context makes the important signal hard to find.

Agent Parallelization

Agent parallelization measures how the person uses multiple AI-agent workstreams when splitting the work would materially help. It includes giving each agent a distinct responsibility, respecting dependencies, communicating changes that affect another workstream, and deciding how the results fit together. Running several processes is not enough. Parallelization becomes a skill when the workstreams have meaningful boundaries and their results are integrated deliberately.

Important boundary

This Skill evaluates human coordination of distinct AI-agent workstreams. Ordinary program concurrency, tool calls, shell processes, and provider requests are not parallel AI-agent workstreams by themselves.

Collaboration

Collaboration measures whether the person and the agent maintain a shared, usable task state. It appears in the feedback that corrects a direction, the answers that resolve a real ambiguity, the decisions that change scope, and the handoffs that let work continue without losing what the person has already decided. Collaboration is not a measure of politeness or message volume. It is the quality of the coordination that lets the next action be taken for the right reason.

Ownership

Ownership measures whether the person remains responsible for consequential judgment while using the agent for speed and reach. It includes setting or deliberately adopting goals, weighing trade-offs, evaluating proposals, challenging mismatches, and calibrating trust to the evidence available. The agent can recommend, investigate, and implement, but the person must retain the authority to decide what the work means and whether it is ready to rely on.

How the Skills fit together

These Skills are related, but they answer different questions. Planning can define a good route without proving that the result works. Verification can find a failure without deciding who should resolve the resulting trade-off. Context operations can preserve the right information while ownership decides whether that information is sufficient to accept the agent's proposal.

Taken together, the Skills describe a practical form of engineering judgment for a semantic programming system. They ask whether a person can direct an agent, give it the right world to work in, inspect what it returns, and remain the author of the decisions that matter.

How Skills Are Scored

The production V2 scorer evaluates the human's contribution to each attribute; it does not treat the agent's implementation quality as human evidence. One complete-conversation call evaluates the four Planning attributes. Each human turn also receives a windowed judgment for the other five Skills, with two preceding turns and at most one following turn supplied for context.

Why Not demonstrated matters

An attribute can be absent because the behavior was not needed or the supplied conversation did not show a human contribution to it. Not demonstrated contributes neither credit nor penalty; it is not a Bad rating.

Evidence Contract

• Attribute records carry a rating, concise rationale, rationale locations, and model provenance.
• Planning is judged once over the complete conversation; the other attributes are judged in human-turn windows.
• The scorer must distinguish human decisions from agent work, quoted text, and automation.
• Missing, truncated, private, ambiguous, or mismatched evidence is excluded or unresolved evidence, not a normal Bad rating.

Serving Attribute Roster

The tables below are the current production V2 serving roster. These are attributes of the human's interaction with the agent, grouped under six Skills.

Planning

Planning captures how the human establishes an actionable direction before and while work proceeds. These four attributes are judged over the complete conversation.

• Sets a clear, actionable goal
• Makes constraints clear
• Sets priorities when goals compete
• Clarifies what the AI can decide

Verification

Verification captures whether the human asks for and responds to checks in a way that matches the material stakes.

• Asks agent to verify that code was written according to specification
• Explains what they checked and observed
• Follows up on important findings
• Makes clear what is checked and what is uncertain

AI Context Operations

AI Context Operations captures whether the human supplies, preserves, and updates the context the agent needs.

• Provides the context the AI needs
• Keeps instructions focused
• Carries useful context into follow-ups
• Fills in important information gaps
• Keeps related work up to date

Agent Parallelization

Agent Parallelization captures how the human coordinates multiple AI agents when the work is split across them.

• Gives each AI agent a clear role
• Keeps dependent tasks aligned
• Flags changes that affect other agents
• Decides how to use each agent's results

Important boundary

This Skill evaluates human coordination of distinct AI-agent workstreams. Ordinary program concurrency, tool calls, shell processes, and provider requests are not parallel AI-agent workstreams by themselves.

Collaboration

Collaboration captures how the human keeps the shared task state actionable as the agent asks questions, receives feedback, and crosses handoffs.

• Gives feedback the AI can act on
• Answers questions needed to move forward
• Makes changes in direction clear
• Shares what the AI needs at handoffs

Ownership

Ownership captures whether the human retains direction, evaluates proposals, and calibrates reliance on evidence.

• Sets the direction of conversation
• Owns important trade-offs
• Explains what matters to the decision
• Evaluates the AI's suggestions
• Challenges suggestions that don't fit
• Redirects the AI when needed
• Bases trust on what is actually supported
• Resolves conflicting information before agreeing

Pipeline Map

AI Skills has three major operating surfaces. The Skill contract sits across them: evidence is collected upstream, scored by the serving release, and interpreted in analysis and product experiences.

Reading a Skill

• A Not demonstrated attribute usually means the behavior was not needed or the conversation did not show a human contribution; it is excluded from the Skill aggregate.
• A beginner Skill result should be traced to Bad or Acceptable attributes and their evidence, not inferred from the aggregate alone.
• An intermediate Skill result reflects adequate behavior with room to improve; inspect the attributes for the evidence.
• An expert Skill result means the human demonstrated the relevant behaviors for the opportunities present.
• A Skill result is meaningful only within its serving release and evidence provenance.

Interpretation rule

Start with the Skill aggregate for orientation, then inspect the attribute ratings, rationales, and bound evidence before drawing an engineering conclusion.

A preview of the six skills, and what an assessment can tell you about a team, is at weaveos.com/ai-skills.