Graph Agentic Coding Extends AI Workflows
Featuring
When an agent works through a multi-step task, the initial prompt quickly becomes only one part of the job. The agent needs to retrieve information, call tools, keep track of earlier decisions, carry state from one step to the next, and check whether the work holds up.
Kewin Wereszczyński and Lech Kalinowski unpacked graph agentic coding as a way to structure that work. The conversation covered how graphs connect the parts of an agent workflow: prompts, actions, tool calls, source material, plans, outputs, and review steps. That structure gives a system a record of what has happened and what it should do next.
From chat to agent workflows
A chat interaction starts with context and ends with an answer. Agent workflows extend that sequence. They can retrieve documentation, search for supplementary information, create a plan, perform actions, review the results, and continue when a task requires another pass.
It’s basically a loop. An agent receives a goal, collects the context it needs, breaks the work into steps, and carries those steps through execution and verification. Coding harnesses already include parts of this pattern through prompts, tool calls, planning, and test runs. A graph makes the relationships between those parts explicit.
State gives agents a way to continue
Longer tasks need a record that survives beyond a single context window. In a graph workflow, state can include the current goal, past outputs, tool results, decisions, files, references, and the next action. Each node holds a piece of the work. The links explain how those pieces relate.
That model also makes failures easier to handle. A workflow can return to an earlier state, retrieve a missing source, repeat a failed step, or hand a task to a separate agent with the relevant context attached. LangGraph was part of the discussion because it provides a way to model these states and transitions directly.
Retrieval is more than a vector database
The conversation also expanded the usual definition of retrieval-augmented generation. Retrieval can involve a local knowledge base, web search, a file system, an internal tool, or a transformation that prepares source material for the next step. Each retrieval step adds context that the agent can use while it works.
This changes how teams can think about context. A coding agent does not need every project detail inside one large prompt. It can retrieve the relevant information when the task reaches the point where that information is needed.
External memory can carry organizational knowledge
The session connected graph agent workflows with the idea of a second brain. Teams already store decisions, notes, policies, documentation, and technical references across Markdown files, Notion, Obsidian, and other knowledge systems. Links and references can turn those separate notes into a connected source of context.
An agent can then navigate that knowledge through the relationships between documents, people, projects, and decisions. Permission boundaries remain part of the design. Access should reflect the information an agent needs for its assigned work.
Planning, execution, and verification can be separate jobs
A workflow can assign different responsibilities to different agents. One agent can prepare a plan. Another can execute a focused task. A verifier can inspect the output, run tests, or decide whether the workflow should continue. An orchestrator can coordinate those steps and pass the right context between them.
That division gives teams more control over work that grows beyond a single agent session. It also gives reviewers clearer points to inspect before an action reaches production systems or changes shared knowledge.
Agent workflows need guardrails
Connected workflows can run for longer and take broader actions, so the session also covered the need for limits. An agent with broad write access can reorganize notes, change references, or alter a knowledge base in ways that damage existing structure.
Stable memory, scoped permissions, verification steps, and review points keep that work under control. Teams can begin with small workflows, observe their behavior, and expand their responsibilities once the system has earned that trust.
