Agent-Coder

Agent-Coder is a decision-critical, token-efficient AI software-development agent for Visual Studio Code. It helps you understand a real codebase, plan changes, investigate failures, review risks, and safely propose edits—without treating an AI response as authority over your files or project state.

It is built for developers who want AI assistance to be useful in day-to-day coding work while retaining control over what is read, changed, executed, and accepted.

What makes Agent-Coder different

Many AI coding tools begin with a prompt and immediately produce code. Agent-Coder starts with the project:

• It indexes the workspace, detects languages, build systems, frameworks, entry points, dependencies, and source layout.
• It gathers relevant evidence from actual files before asking an AI provider to reason about the task.
• It validates structured AI responses, checks project rules, presents a real diff, and normally requires your approval before a write reaches disk.
• It keeps deterministic local systems responsible for project state, file integrity, policy enforcement, validation, and completion status.

In short: the AI can recommend and generate; Agent-Coder checks, constrains, and coordinates.

How it benefits AI-assisted developers

Agent-Coder is designed to make AI coding more dependable and less wasteful.

A practical workflow

1. Open a workspace and let Agent-Coder analyse it. A startup scan warms the project index; you can also run Agent-Coder: Analyse Workspace at any time.
2. Ask, search, plan, or analyse before editing. Use ordinary chat language or focused commands such as /plan, /debug, /review, /design, and /security.
3. Request a change. Agent-Coder selects grounded context and creates a patch proposal rather than directly overwriting files.
4. Inspect the real diff. Review validation findings, rule/invariant results, and the exact proposed change.
5. Approve or reject. By default, a proposal is not applied until you explicitly approve it. Applied edits can be rolled back.
6. Build, test, and validate. Use the detected project toolchain and retain the evidence needed to judge whether the task is actually complete.

This workflow deliberately makes the human decision point visible. You can use Agent-Coder for explanation and analysis only, or allow it to help progress through carefully bounded repairs and edits.

Core capabilities

Grounded project understanding

• Workspace discovery, incremental indexing, project snapshots, language/build/framework detection, and generated-code suppression.
• Hybrid and semantic code search, symbol slicing, context manifests, progressive disclosure, and relevance scoring.
• Architecture knowledge graph, dependency visualiser, project map, call-graph analysis, concurrency analysis, and resource-ownership analysis.
• Language adapters for 27 languages, including TypeScript/JavaScript, Python, C/C++, C#, Go, Rust, Swift, Delphi, Lua, assembly, and web assets.

AI-assisted reasoning

• Task decomposition and requirements extraction.
• Decision analysis with evidence, assumptions, alternatives, and confidence scoring.
• Architecture, debugging, refactoring, performance, security, comparison, research, and teaching modes.
• Debug hypotheses, diagnostic-artifact analysis, experiment tracking, and root-cause verification.
• Test generation and test-adequacy assessment, with structured-response validation before results are used.

Safe editing and validation

• File hashes and guarded paths help prevent a patch from being applied to stale or protected content.
• Patch schemas, semantic checks, patch-granularity checks, quality rules, project invariants, and prompt-injection detection run before an edit is offered.
• Diffs are previewed before approval; backups and rollback support are available after application.
• Multi-file proposals, generated tests, documentation comments, and deterministic formatting cleanup use the same review-and-approval model.
• Build, test, validate, repair, regression, baseline, performance-budget, schema, API-contract, serialisation, and accessibility workflows help verify outcomes.

Engineering governance and learning

• Architecture decision records, project invariants, user preferences, technical-debt tracking, provenance, solution patterns, and documentation-impact analysis.
• Persistent project memory, rejected-approach tracking, corrections, checkpoints, and diagnostic-causality reduction.
• Git status/diff support, checkpoints, branch and commit-message assistance, plus change-impact awareness.
• Controlled autonomy limits for time, files, commands, repair attempts, token use, and AI requests.

Visibility and optimisation

• Dashboards for project mapping, dependencies, technical debt, project score, AI confidence, token use, memory timeline, workspace/build/memory optimisation, and engineering control.
• Token policies, cost/benefit gates, response caching, resource-pressure handling, model capability routing, and performance acceptance checks.
• Workspace goals, design proposals, refactoring simulation, performance simulation, learning-mode symbol explanations, and AI pair-programmer diagnostics.

Choose how you interact with it

Agent-Coder provides several interfaces over the same controller and safety pipeline:

• Agent-Coder sidebar: a persistent conversation view in the VS Code Activity Bar.
• Native VS Code Chat: use @agent-coder with commands such as /ask, /plan, /fix, /validate, /callgraph, /debug, /review, and /teach.
• Command Palette: use focused commands for workspace analysis, build/test/repair, governance, project analysis, settings, and dashboards.
• MCP bridge: use the standalone agent-coder-mcp server from compatible tools such as Claude Code or Codex CLI while an Agent-Coder-enabled VS Code workspace is open.

All of these routes share the same workspace understanding, proposals, approvals, and controller logic—there is no separate “unsafe chat” path.

AI providers and privacy

Choose the provider that suits the task and your environment:

• Anthropic Claude
• OpenAI / ChatGPT
• OpenAI Codex models
• DeepSeek
• A local OpenAI-compatible server such as Ollama, LM Studio, llama.cpp, text-generation-webui, or vLLM

Provider keys entered through the extension are stored in VS Code's encrypted SecretStorage, not in project configuration files. Local-server settings are stored as machine-local preferences. Agent-Coder also includes privacy and network-access policies so projects can restrict vendor-provider access when required.

Lite and Pro

Agent-Coder is available in two tiers:

• Lite: free access to the foundation through security and trust-boundary capabilities (Phases 1–20).
• Pro: adds toolchain and compatibility awareness, Git safety, advanced review/economy features, project memory, debugging intelligence, validation, provenance, plugins, autonomy controls, and the advanced dashboards (Phases 21–40).

The UI identifies locked features and provides licence activation and status commands. An active licence is stored securely through VS Code rather than in the workspace.

AGENT-Coder by Daniel J. Hobson of Ultimanium Designs 2026-2027
Copyright (C) by Daniel J. Hobson, Australia, 2026-2027
Website: https://ultimanium.com/
License Type: Lite/Pro