Agent runtime, Software architecture

Titan Code

From a task to verified work.

A coding agent that works across files, browsers and desktop applications, with one execution loop keeping actions, observations and session history connected.

One loop. Many surfaces.

  1. TaskThe instruction
  2. Agent loopPlan, act, observe
  3. Policy & consentCheck before execution
  4. Tool executionFiles, shell, browser, desktop, MCP
  5. Observed resultWhat actually changed
  6. Model providersA shared stream contract
  7. Session contextSQLite, history, memory
Tool results return to the agent for the next decision. Provider adapters and session context support the same loop.
  • Task connects to Agent loop.
  • Agent loop connects to Policy & consent.
  • Policy & consent connects to Tool execution.
  • Tool execution connects to Observed result.
  • Observed result connects to Agent loop.
  • Model providers exchanges context with Agent loop.
  • Session context exchanges context with Agent loop.

From a task to verified work

Titan Code carries a task from understanding a codebase to taking action and checking what changed. Model providers share the same agent loop, while tools connect it to files, shell commands, browser pages, the Windows desktop and MCP servers.

Every tool call is checked against its schema and permissions before execution. Observed results return to the conversation; SQLite sessions and a redacted journal preserve the work across interruptions. A terminal interface makes progress, waiting and approval visible while the task is running.

01

Understand the surface

Semantic page snapshots provide named controls and generation-safe references. Screenshots cover visual state and native applications.

02

Keep action accountable

Schema checks, permissions and scoped confirmations precede tool execution. The journal records calls and outcomes with sensitive values redacted.

03

Continue with context

Resumable sessions, repeat guards and an optional typed-decision judge support longer tasks. The judge advises; deterministic code retains the hard-stop rules.

A boundary around action

Shell execution passes through separate checks. Command analysis identifies the operation; path resolution follows its real destination; supported operating systems enforce filesystem confinement. The runtime reports when a platform has no confinement backend.

  1. 01

    Consent

    Review the action where required.

  2. 02

    Command analysis

    Inspect syntax, wrappers and writes.

  3. 03

    Path containment

    Resolve the effective destination.

  4. 04

    OS confinement

    Enforce the boundary where supported.

Explore the engineering

Technical documentation and a runnable reference core.

GITHUB