@pallavishekhar_: AI Coding Agent SWE - High-Level Architecture This shows the main pieces of an autonomous coding agent and how a GitHub…

AI Coding Agent SWE - High-Level Architecture

This shows the main pieces of an autonomous coding agent and how a GitHub issue flows from a filed ticket all the way to a pull request a human reviews.

• GitHub issue webhook: This is where a run starts, a real GitHub issue from a developer, sometimes a detailed one with a stack trace and sometimes just one line like “login is broken on mobile”.

• Ingestion and triage: This receives the issue, decides whether the ticket is even actionable, and creates a run with its own budget so we never spend without a limit.

• Agent orchestrator: This is the brain of the system, the autonomous loop that talks to the model, decides which tool to call next, applies the edits, runs the tests, and feeds the results back.

• Model gateway: A thin routing layer in front of the models, so we can send hard reasoning to a strong model and easy steps to a cheap model.

• Tools: These are the agent’s hands, with code search, a file reader, a file editor that applies diffs, and a command runner to build and run tests.

• Code intelligence: This builds a code map and a symbol index of the repo, so the agent can ask “where is this function defined” and get an exact answer, like a developer using go-to-definition.

• Sandbox environment: An isolated place with a fresh checkout of the repo at one commit where the agent builds and runs tests, with no production access, no secrets, and no open network.

• PR builder: When the tests are green and the issue is fixed, this turns the final diff into a pull request with a clear description for a human to review.

• Guardrails: These enforce the hard limits, a cost cap, an iteration cap, and a no force-push rule, so an autonomous loop cannot run forever or burn money.

• Observability: This traces every tool call, every model response, every test result, and the cost, so we can replay exactly what the agent saw and did.

• Math behind Attention- Q, K, and V

• Math behind √dₖ Scaling Factor in Attention

• Math Behind Backpropagation

• Math Behind Gradient Descent

• Math Behind Cross-Entropy Loss

• Math Behind RoPE (Rotary Position Embedding)

• RMSNorm (Root Mean Square Layer Normalization)