The Coding Harness (CodingAgent)

CodingAgent is a thin harness — shipped in adk-agent behind the coding feature — that wires the dev tools, a planning write_todos tool, and a minimal coding system prompt onto a normal LlmAgent. It's configuration over LlmAgent, not a new agent type, so it keeps the default adk-agent build lean (the feature pulls in adk-devtools).

Build one

use adk_agent::coding::CodingAgent;
use adk_devtools::Workspace;

let coding = CodingAgent::builder()
    .model(model)                              // any adk-model provider (required)
    .workspace(Workspace::new("./my-repo"))    // sandboxed (required)
    .instruction("Follow the project's existing style; prefer small diffs.") // optional, appended
    .build()?;

let agent = coding.agent();   // Arc<dyn Agent> — hand to a Runner

Builder options:

What's wired

• The dev toolset — read/write/edit/glob/grep/bash, scoped to the workspace.

• write_todos — a planning tool the model uses to record and update a short task list. Read it back from the harness:

  for todo in coding.todos() {
      println!("[{}] {}", todo.status, todo.content);  // pending | in_progress | completed
  }

• A minimal prompt — sub-1k-token base instructions ("explore before you change", "read before you edit", "verify by running tests", "track a plan"), kept small on purpose so capabilities come from tools and skills rather than a huge prompt.

The loop

You run a CodingAgent like any agent — through a Runner. Within a turn the underlying LlmAgent already executes a plan → act → observe loop: it calls tools, sees their results, and continues until it's done. A typical turn:

write_todos(...)            # plan
glob / grep / read_file     # explore
edit_file / write_file      # change
bash("…test…")              # verify
write_todos(... completed)  # update plan
→ final summary

For multi-turn work, reuse the same Runner + session across calls and the agent builds on its own prior work (it re-reads and edits the files it wrote). See the coding_agent example (multiturn mode).

Memory & skills

Because it's an LlmAgent underneath, everything else in ADK composes:

• Attach a MemoryService (e.g. the bi-temporal knowledge graph) on the Runner for cross-session project memory.
• Add lazy skills via adk-skill to extend capabilities without bloating the prompt.
• Hooks/guardrails (adk-plugin, adk-guardrail) gate tool calls deterministically.

Running it

The harness produces an Arc<dyn Agent>; drive it with adk-runner (see the quick start) or use the CLI, which wraps all of this.

Next: The CLI →