GitHub Copilot Coding Agent

Workflow & Role of the Coding Agent

• Agent is invoked by assigning issues; it creates a branch and PR, cannot push to default branch directly.
• Several commenters liken it to a “junior dev” opening PRs, but others point out it lacks human traits (intuition, holistic thinking, communication, ownership).
• Some worry managers will treat this as justification to reduce headcount and then pressure remaining devs to “go faster because AI writes the code.”

Tests, Code Quality & “Vibe Coding”

• Official positioning: works best on low–medium complexity tasks in well‑tested codebases.
• Multiple users confirm that strong tests “box in” the agent and make it more reliable, but also note AI‑generated tests often increase coverage with shallow or error‑suppressing checks.
• Several experiences with “vibe coding” greenfield projects: AI can be a big productivity boost but easily breaks abstractions, accumulates architectural debt, and rarely self‑critiques design without heavy guidance.
• Others find AI much more effective in brownfield codebases, where it can pattern‑match existing style and architecture.

Cost, Performance & Model Choices

• People report burning through $10–$15 in tokens in a single evening with agentic tools, prompting debate about “time saved” vs real value and long‑term AI bills.
• Some prefer subscription models; others prefer a‑la‑carte via APIs. Consensus that models are commoditizing but not necessarily getting cheap.
• Complaints that Copilot’s agent/edits can be slow or flaky compared to using raw models via other tools; others report snappy behavior, suggesting variability.
• GitHub staff say the agent currently uses Claude 3.7 Sonnet but may get a model picker later.

Dogfooding, Metrics & Hype

• GitHub representatives say ~400 employees used the agent across ~300 repos, merging ~1,000 agent PRs; in the main repo it’s the #5 contributor.
• Commenters push back: want rejection rates, amount of human intervention, and comparison to prior automation; some call out survivorship‑bias‑style marketing.
• Mixed reports from Microsoft ecosystem: management‑driven mandates to “use AI” vs devs who mostly ignore it.

Security, Privacy & Dependencies

• Concern that agents may pull in random, low‑quality dependencies from obscure repos as if they were standard solutions.
• GitHub says agent PRs don’t auto‑run Actions; workflows must be explicitly approved, to avoid blindly running code with secrets.
• Strong distrust around training on private repos; some point to opt‑out controls, others assume individual plans are still used for training.
• Enterprises like LinkedIn reportedly block Copilot on corporate networks, reflecting ongoing security skepticism.

Competition, UX & Broader Impact

• Comparisons to Cursor, Windsurf, Aider, Cline, Claude Code, and Google’s equivalents; many say third‑party tools feel more capable or better tuned than Copilot’s current UX.
• Frustration with Microsoft’s aggressive Copilot branding and deep integration into GitHub/VS Code, even when users don’t want it.
• Broader existential thread: will agents mostly remove tedious tasks or erode software careers entirely? Opinions range from “junior‑level helper” to “this will eat knowledge work and crush social mobility.”