how coding agents got here
this timeline follows several developments that help explain today’s coding agents: models that work with code, loops that call tools, environments that run those tools, and interfaces for reviewing the result. the dates mark the linked paper or announcement. the broader history includes overlapping work and many other contributors.
-
the transformer creates the foundation
Attention Is All You Need introduced the Transformer, an attention based architecture designed for parallel training. later large language models built on that architecture.
the practical connection is sequence modeling: instructions, code, and later tool results can become context for another prediction. the architecture gives part of the foundation; the surrounding tool system supplies execution.
-
prompts become a general interface
GPT-3 showed how scaling improved performance on tasks specified through text and examples, using the same model without a separate gradient update for each task.
programming a model through language started to look like a general interface. reliable action loops came later.
-
code generation reaches the editor
OpenAI published Codex, a GPT model trained on public code, while GitHub introduced Copilot as an editor product powered by it. the Codex paper also introduced HumanEval as a way to measure code generation from docstrings.
generation moved from a research result into the place many developers already worked.
-
the model gets an action loop
ReAct described an interleaved loop of reasoning and actions against external systems.
the core pattern now appears across agent systems: inspect state, decide what to do, use a tool, read the result, and continue.
-
the terminal and editor become agent surfaces
aider’s early release, published on June 8, 2023, supported repository editing, diffs, git commits, and undo in a terminal workflow. Cursor was building around codebase context, inline edits, and constrained agents inside its editor.
these tools made the surrounding interface part of the value. context selection, change review, and recovery became product decisions.
-
the benchmark becomes a real repository
SWE-bench collected 2,294 real GitHub issues and their pull requests from twelve Python repositories. the paper reported that its best evaluated model, Claude 2, resolved 1.96 percent of those issues.
that historical result describes the paper’s evaluation. the benchmark gave the field a harder target: repository work required long context, execution environments, and coordinated edits across files.
-
Devin demonstrates a software task
Cognition introduced Devin with a demonstration of an agent using a shell, editor, and browser. the launch brought the idea of delegating a whole software task into a single visible workflow. the video records Cognition’s demonstration at the time.
-
the task becomes the unit of work
GitHub previewed Copilot Workspace as a task based environment that could move from an issue or prompt through planning, implementation, testing, and execution.
the interface moved above individual completions and toward a reviewable sequence of agent work.
-
the harness becomes its own engineering problem
SWE-agent showed that the interface between a model and a computer materially affects results. its custom interface let the agent navigate repositories, edit files, and run tests.
this made the harness legible as its own engineering layer alongside model capability.
-
Claude uses the screen
Anthropic introduced computer use in beta: Claude could interpret screenshots and act through mouse and keyboard commands. this October 2024 demo shows the screen itself becoming an interface for agent actions.
-
tools get a shared protocol
Anthropic released the Model Context Protocol as an open standard for connecting assistants to tools and data sources. the launch included a specification, software development kits, local server support, and an open repository of servers.
agent capability started depending more visibly on the connections around the model.
-
terminal and cloud agents spread
GitHub introduced agent mode for Copilot in VS Code in February. Anthropic launched Claude Code as a terminal research preview later that month. OpenAI launched Codex as a cloud software engineering agent in May.
the product decision now included execution location, permissions, context, and review surface.
Anthropic’s February 2025 launch recording shows that workflow in the terminal.
-
supervision gets a dedicated surface
on February 2, OpenAI launched the Codex app for supervising agents across projects, threads, and isolated worktrees. on April 2, Cursor introduced a workspace centered on parallel local and cloud agents, review, and handoffs.
supervising several tasks puts scope, context, permissions, and integration into the interface. each completed task still needs an understandable result and a place for review.
what the timeline explains
the useful pattern is how much of the work moved into the loop. a completion suggests code; a repository task can inspect files, run a command, use its output, and return a change for review. improvements can come from the model, the tools it receives, or the environment and interface around them.
that gives a practical way to evaluate a new release. look for the part of the workflow it changes, then try a task that exercises that part. a new model score, a different tool interface, and a better review surface each need an example suited to the claim. the product guides carry the current details; the timeline provides the context for understanding what changed.
