How AutoDev Turns Prompts into Custom Local AI Coding Agents

Limitations of existing AI coding assistants

Current tools such as Copilot, Cursor, and AutoGPT‑style projects act mainly as advanced autocomplete engines. They lack structured task understanding, contextual awareness, and the ability to orchestrate multi‑step actions based on developer intent.

AutoDev paradigm: Prompt as Agent, Configuration as Capability Chain

Core principles

Agent structure is declarative, capabilities are composable, behavior is extensible.

A single prompt plus a concise configuration yields a fully private, locally runnable coding agent.

Developer‑driven capabilities

Define custom ability chains (e.g., read a file → retrieve related functions → generate tests → commit).

Inject personal prompt styles, sub‑task planning logic, or plug‑in modules.

Select any local large language model (e.g., Deepseek, Qwen, GLM) for inference.

Configure interaction modes for IDE context menus, command‑line interfaces, or @‑mention actions.

Example 1: One‑minute agent declaration for Protobuf generation

The following YAML creates an agent that, when a user selects a requirement, generates a Protobuf IDL via a right‑click menu.

Execute fetch-teamai-spec.s to obtain the specification.

Inject the fetched context and combine it with the user’s selected requirement.

Automatically generate a .proto file that matches the spec and open it for editing.

Example 2: Orchestrated database‑design agent

---
name: "Design Database"
variables:
  "story": /any/ thread(".devin/shell/github-issue.curl.sh") | jsonpath("$.body")
afterStreaming:
  case condition {
    default { execute "gen-sql.devin", $story, $output }
  }
---
You are a professional DB administrator. Based on user needs, select the best tables.
---
User database: $databaseInfo
- User tables: $tables
---
Example request: Compute average trip length per user type using tables trips, users, Subscriber_type.
---
Please return the selected tables as a list without explanation.

This configuration produces an agent that parses a GitHub issue to extract a user story, selects appropriate database tables, generates SQL via the gen-sql.devin agent, and can pass the query results downstream.

Example 3: In‑chat code generation

Setting agentic: true and mapping $selection to $input lets the agent receive chat input and generate the requested code automatically.

---
name: "Design Protobuf IDL"
actionLocation: ContextMenu
variables:
  "spec": /any/ thread(".devin/shell/fetch-teamai-spec.sh") | jsonpath("$.[1].content")
onStreamingEnd: { parseCode | saveFile | openFile }
agentic: true
---
Please generate a Proto file according to the user’s input and the provided spec.

Conclusion: a developer‑driven agent workshop

Developers define agent behavior instead of accepting preset actions.

Modules can be composed and updated independently.

Model choice is under developer control, avoiding platform lock‑in.

Project repository: https://github.com/unit-mesh/autodev

Documentation & quick‑start guide: https://ide.unitmesh.cc/docs