GitHub’s 2025 roundup spotlights three forces reshaping developer workflows: agentic AI, the Model Context Protocol (MCP), and spec‑driven development. Here’s how to put them to work now.
Source: GitHub Blog roundup.
The signal
AI is moving from clever demos to dependable systems. The common thread: standardize tool use, constrain outputs with specs, and measure results like any other service.
- Agentic AI: multi‑step planners that call tools, retrieve data, and self‑check before returning results.
- MCP: a common protocol so different models and editors can use the same tools with less glue code. See modelcontextprotocol.io.
- Spec‑driven development: define goals, inputs, constraints, and output schemas to reduce hallucinations and speed reviews.
What to do now
- Standardize tool calling: pick or build MCP‑compatible tools so your agents and IDEs speak the same language across vendors.
- Write task specs: capture intent, inputs, constraints (latency, cost, PII), and an output JSON schema before you prompt.
- Add evals early: create small, repeatable checks (accuracy, safety, latency) and run them on every change. A starting point: OpenAI Evals.
MCP quickstart checklist
- Inventory the tools your agents need (search, CRUD APIs, vector DB, billing).
- Wrap each as an MCP server with clear input/output JSON schemas and auth.
- Adopt an MCP‑capable client (editor extension, CLI, or agent runtime) to wire tools without custom adapters.
- Log all tool calls and responses; include model, prompt, and context for tracing.
- Gate deployments with evals that cover happy paths, edge cases, and failure handling.
Spec‑driven prompts in practice
Before writing a prompt, write the spec. Then make the prompt reference the spec explicitly so the model is constrained.
- Goal: what success looks like in one sentence.
- Inputs: allowed fields and types.
- Constraints: policies, time/cost budgets, data sensitivity rules.
- Output schema: a JSON structure the service must return.
- Acceptance test: a minimal eval that passes only if the output matches the schema and business rule.
Agentic AI without the hype
Start simple and graduate to agents only when needed. Use agents when a task requires multiple tool calls, branching, retries, or long‑running workflows; otherwise, prefer single‑shot functions with strict schemas.
- Great fits: data enrichment pipelines, code refactors with tests, customer support workflows with retrieval and ticketing.
- Poor fits: fixed reports with deterministic SQL, safety‑critical steps without robust evals.
Takeaway
Treat LLMs like any service: standardize tools (MCP), constrain outputs (specs), and measure quality (evals). That’s how you move from prototype to production.
Get smarter about AI, weekly
Enjoy concise, practical AI insights like this? Subscribe to The AI Nuggets newsletter: theainuggets.com/newsletter
