LLM ORCHESTRATOR

Overview 01 Plan First 02 Route Intelligently 03 Execute & Learn

Prompt Optimization LayerTune prompts so cheaper models match production quality.Hybrid Model ArchitectureExpensive orchestrator delegates to cheap sub-agents. No wasted spend on simple tasks.Smart Model SelectionAuto-picks the best model per prompt, using expensive models only when needed.Model Limitation AwarenessKnows model limits — escalates or reroutes instead of hallucinating.

Granular Plan ModeGenerates detailed step-by-step plans when needed, and skips planning for simple tasks.Pre-Execution Token EstimationSee token usage and projected cost before build commands run.Token Budgeting SystemSet token budgets per session, task, or team — enforced before compute runs.Structured Plan OutputOutputs interactive plan.html — reviewable decisions, not markdown walls.

Smart Context ManagementOnly pass context relevant to the latest prompt — not the full session history.Context & Memory ControlAdd or remove context and memory to control what the model sees.

Interactive Think ModePause, redirect, or correct the agent mid-run — no full restart.Skill BuilderBuild skills with test and evaluation modes to benchmark and fine-tune behavior.CLI + Local Web IDEProvides a terminal CLI and localhost web IDE with embedded chat.AI Collaboration ModeShared agent sessions and handoffs — break out of solo AI silos.

Cross-User LearningIndexes resolved failures org-wide, like Stack Overflow for AI agent runs.Automatic Repository RulesGenerates repo rules from your codebase patterns on init.

How it works overview

Phase 03

Execute & Learn

Run once. Route smarter next time.

Overview

Execution closes the loop: Devtor runs the approved plan via CLI or local web IDE, streams progress in the terminal, and lets you interrupt with think mode. Smart context management keeps only relevant messages in each call.

In depth

After each task, patterns are cached — task shape, files touched, model used, tokens consumed. Cross-user learning indexes resolved failures so similar challenges retrieve proven fixes. Collaboration mode lets teams share sessions instead of working in AI silos.

What happens

1Execute against the approved plan and scoped context
2Stream intermediate steps; allow think-mode interrupts
3Record actual tokens, latency, and cost vs. estimate
4Cache routing patterns for similar tasks
5Index resolutions for cross-user retrieval

Terminal output

terminal

✓ Completed in 3.2s

✓ Tokens used: 1,847 / 2,400 · budget remaining: 62%

✓ Cost: $0.006

✓ Pattern cached · fix indexed for org

✓ 12 similar tasks will benefit

Benefits

Interactive think mode for mid-run corrections
CLI + local web IDE for flexible workflows
Smart context — only relevant messages per call
Cross-user learning from shared resolutions

Related capabilities

Prompt Optimization Layer

Tune and augment prompts so cheaper models perform like expensive ones.

Hybrid Model Architecture

One super agent. Many cheap sub-agents. Automatically.

Interactive Think Mode

Interrupt an agent mid-execution. Redirect. Continue.

Build, test, and benchmark custom agent behaviors.

AI Collaboration Mode

Break out of solo AI silos.

CLI + Local Web IDE

Terminal and localhost web IDE — one seamless workflow.

Cross-User Learning

When one instance solves a problem, all instances benefit.

Model Limitation Awareness

Knows what models cannot do — and says so.

Automatic Repository Rules

Coding conventions extracted from your codebase.

PreviousRoute Intelligently

Ready to orchestrate?

Installation guide and CLI — coming soon.