Engineering

Agentic coding assistant for engineers

An in-IDE and CI coding agent that reads the repo, writes and edits code, runs tests, and opens PRs - iterating until the task is green.

Agentic coding assistant: $1.35/run, $67.5k/yr, 57% invisible in accumulated-diff history, agent-loop hops, and full-repo re-embedding. Firedog instruments every iteration across the engineering team.

Model 28%Prompt 16%Orchestration 10%Retrieval 13%History 33%

ComponentWhy it is unattributedcalls$ / run

Model · 28%Code generation and editing (Claude Sonnet 4)Each iteration sends the relevant repo files, the task, and the current diff to Sonnet 4 and receives edited code or a next-step plan.tokens in: 90,000 · tokens out: 7,000Visible on the invoice.10$0.38

Prompt · 16%Coding-agent system prompt and repo conventionsA 7,000-token prefix carrying the agent instructions, repo conventions, and tool schemas is resent on every iteration without prompt caching.tokens in: 70,000 · tokens out: 0Visible on the invoice.10$0.21

Orchestration · 10%Agent loop: plan, edit, test-run, PR hopsEach task runs a planner, file-editor, test-runner, and PR-writer as separate tool calls against GitHub/GitLab and CI. A 10-step task produces ~24 billed hops.tokens in: 25,000 · tokens out: 4,000The plan-edit-test-PR hops are abstracted into one 'agent task' in the tool's logs; the underlying billed calls are never broken out per step.24$0.14

Retrieval · 13%Codebase semantic search and re-embeddingThe agent runs semantic search over the repo on each iteration and re-embeds changed files - and often the whole repo - rather than updating incrementally.tokens in: 40,000 · tokens out: 0Codebase embedding runs on the embedding-provider invoice; full-repo re-embeds are a default that never surfaces in the engineering LLM spend view.20$0.18

History · 33%Conversation and accumulated diffs re-sent each iterationThe full agent conversation and every accumulated diff is resent on each iteration so the model retains prior reasoning. By step 10 this is 120k+ tokens re-billed each call.tokens in: 150,000 · tokens out: 0Accumulated agent state is framed as 'context window' by the coding tool, not as re-billed tokens, so it never appears in per-call spend reporting.10$0.45

An agentic coding task costs ~$1.35; run 8 a day across 25 engineers and it is ~$67.5k/yr - 57% of it accumulated-diff re-injection, agent-loop hops, and full-repo re-embedding no one instrumented.

$1.35

cost per run

$68k

per year

57%

unattributed today

$10.80

per user, per day

Scenario: ~25-engineer team, agentic coding assistant. 25 active users, 8 runs/day, 250 working days. 10 documents/run. Avg doc: 3,000 tokens.

Default model: claude-sonnet-4

Trigger: Per engineering task or ticket, plus inline completions throughout the day. A developer runs 6-12 agentic tasks daily; each task iterates 6-15 steps.

Who runs it: The engineering team building the product. Each developer runs the agent on tickets from Jira/Linear; it works against the GitHub/GitLab repo and posts results to Slack. Agentic loops and repo context make it the most expensive workflow per run.

What Firedog shows

Token cost per agent iteration, showing the step where accumulated-diff history overtakes actual code generation.
Which tasks trigger full-repo re-embedding versus incremental updates, and what that costs per developer.
Projected annual saving from capping iteration depth with summary compression and caching the agent system prompt.

Decisions this informs

Compress agent history into a rolling summary after step 5 to stop unbounded diff and conversation accumulation.
Cache the 7,000-token coding-agent system prompt and repo-conventions prefix; at 10 iterations/run this is a large recurring saving.
Switch codebase retrieval to incremental embedding so only files changed since the last run are re-embedded, not the whole repo.

Agentic coding assistant for engineers

What Firedog shows

Decisions this informs

Understand every AI dollar.