GitHub CopilotBuild Guild

September 16 meeting notes

Meeting: State of CA GitHub Copilot Build Guild — Launch Session
Topic: GitHub Copilot Token Optimization: Getting More from Your AI Usage
Presenters: Aparna Jain (host), Jason Cook, Kenedy Thorne

Generated by AI from the meeting recording. Be sure to check for accuracy.

Build Guild community launch

Aparna launched the California State Government GitHub Copilot Build Guild, describing it as a statewide community with three goals: learn, share, and build together — an exchange of practical use cases, lessons learned, and challenges, rather than a one-way vendor presentation. Future sessions will be shaped by participant feedback via a topic poll.

• Monthly format: 30 minutes of featured topic & demo, 15 minutes for a community member to show what they built, and 15 minutes for open Q&A. Participants were encouraged to ask questions throughout rather than waiting for the final segment.
• Audience & topics: The Guild is for more than application developers — planned subject areas include Power Platform, data & database development, SQL, Python, notebooks, legacy application modernization, PowerShell, infrastructure as code, QA, and testing.

Token optimization principles

Jason Cook explained that improving agent quality and reducing token cost are linked: choose models deliberately, provide only relevant context, split complex work into stages, and use deterministic controls to reduce rework.

• Quality over counting: Focus on making every token count rather than simply minimizing token numbers — inaccurate or unclear intermediate results compound across multi-step workflows, so even high per-step accuracy can produce much lower end-to-end accuracy over many steps.
• Model selection: Match model capability to task complexity. Higher-reasoning models suit research, planning, architecture, and hard debugging; mid-tier models suit implementation; lighter models suit repetitive edits, documentation, and routine work. Use Auto model selection as a default while learning which models fit which tasks.
• Context management: Provide enough context to solve the task, but no more — narrow context to relevant files, modules, functions, or specs rather than sending an entire codebase.
• Session & task boundaries: Start a new session after completing a feature or checkpoint, clear unrelated context, and break large work into smaller tasks instead of one extensive task list.
• Deterministic guardrails: Add unit tests, linters, architecture conventions, and security scans to agent workflows for consistent early feedback and fewer repeated reasoning/correction loops.

Prompting & agent workflow design

• Prompt construction: Use concise descriptions, imperative goals, and known context — identify the relevant class or function, describe triggering conditions and observed behavior, and specify the intended outcome. Add stop conditions requiring Copilot to ask before taking an uncertain path.
• Staged execution: Separate complex work into research, planning, and implementation phases — a reasoning model investigates and plans, then a lower-cost coding model implements the specified changes.
• Agent configurations: Persistent instructions should contain only small, stable, non-negotiable rules (they're injected every turn); custom agents define task-specific personas; skills encode repeatable actions; MCP tools retrieve dynamic external context.
• Organization standards: Kenedy explained reusable agents can live at the org level for use across projects — keep org-wide instructions general (broad compliance) and project-specific testing rules/skills scoped to the relevant repo.
• Agent composition: Agents can call other agents, letting an orchestrator delegate testing, documentation, user stories, or commits — each with its own assigned model or model range.
• Learning from usage: An "agentic flywheel" — inspect incorrect or inefficient results, identify what guidance would have prevented the issue, and add that learning to instructions, specs, or skills for reuse.

Practical cost reduction techniques

• Scripts instead of AI: Use ordinary scripts for deterministic, repeatable data prep rather than spending tokens on tasks code can do directly.
• CLI versus MCP: Where a CLI or API can perform the same action, it may reduce static context and tool-call overhead compared to an MCP tool manifest.
• Concise outputs: Ask Copilot to return only what's needed (e.g., file names, not full paths/timestamps/permissions) — output tokens are often more expensive than input context.
• Usage analysis: Use Chronicle to review behavior across sessions and identify repeated instructions that should become persistent guidance; inspect context/usage in Copilot surfaces.
• Targeted specifications: For large projects, create per-module spec files documenting purpose, invariants, and dependencies so Copilot can read specs instead of loading the whole codebase.

Orchestration cost controls

In response to a question about agent fleets, Kenedy and Jason explained that orchestration cost depends on task complexity, model choices, and the number of delegated agents. A practical pattern: use a stronger model to create a detailed plan and lighter models to execute subtasks, with a constrained (not indeterminate) number of agents. Auto model selection also provides a 10% discount on the model it chooses. Jason referenced a simple example — the "Ultra Light Orchestrator" — to be shared afterward.

Context, memory & data protection

• Compaction: Compaction reduces conversational history to the important information for later turns and doesn't itself explicitly cost tokens — but frequent compaction is a signal to split work, start new sessions, or introduce orchestration.
• Context ownership: Copilot manages the context window internally (behavior varies by model); users can manage it by clearing or starting a new session at a logical checkpoint. Session streaming is available in public preview via an enterprise admin setting.
• Documentation volume: Every word supplied to the model consumes tokens — normal code documentation is fine, but documentation that greatly exceeds the code it describes is likely excessive.
• Data handling: GitHub Copilot does not use prompts/data to train backend model providers and does not retain those records under applicable protections. For government use, use GitHub Copilot Business or Enterprise — the same guarantees do not apply to Free, Pro, Pro+, or Max plans.

What's next

• Recording and slides shared with attendees (see links above).
• A poll will go out to vote on topics for upcoming sessions.
• Strong interest in a future session on GitHub Copilot agents — tools, models, and orchestration patterns.
• Next monthly session: October 21, 2026 — Agentic Development with GitHub Copilot.