An auto-adaptive
coding-agent harness.
agntro classifies each task and scales its pipeline depth to match — from a
single search → summarize for a research question, to a full
multi-model plan → code → review ladder for a complex feature.
Built on Elixir/BEAM for supervision, crash recovery, and concurrent agents. MCP-native codebase understanding. A mission-control cockpit that streams every agent turn, tool call, and DAG node live over WebSocket — and lets you steer mid-run.
the pipeline scales with the task
trivial
Research question / audit
A question about the codebase, not a change to it.
search graph → summarize · ~3 calls
simple
Small, localized change
A single edit with a drift check.
plan → code → drift review · ~10 calls
medium
Feature with a contract
Spec extraction anchors the work; a contract-verifier runs the tests.
spec → plan → test → code → contract-verify · ~25 calls
complex
Multi-component feature
A planning ladder with variations, delegated sub-tasks, and the full review panel.
spec → plan-ladder → delegate → test → code → full review · ~100 calls
{ adaptive }
Auto-adaptive routing
Every task is classified before any work begins. A research question
doesn't trigger a coding pipeline; a complex feature doesn't get a
single-pass plan.
{ mcp }
Codebase-grounded
Code understanding comes from the codebase-memory graph — structural
search, call paths, architecture — not from re-reading files every turn.
{ beam }
Built on Elixir/OTP
Supervision trees, crash recovery with checkpointed state, and
concurrent agent processes. The harness survives what kills a script.
{ cockpit }
Mission-control UI
A real-time cockpit: the task DAG draws itself as the planner
decomposes, agent turns and tool calls stream over WebSocket, and
you can steer mid-run. Per-agent swimlanes, convergence ladders,
and full conversation transcripts with reasoning blocks.
{ waveplan }
Hardware-aware scheduling
An analytical scheduler predicts the decode ceiling per model per
GPU — before you run anything. The predicted-vs-measured overlay
shows theory against reality, and VRAM-fit analysis tells you what
fits and what doesn't.
the cockpit
live
Real-time pipeline visibility
Every agent turn, tool call, and LLM telemetry event streams over
WebSocket the moment it happens. Phase ribbons, the synthesized
task DAG, and a converging fix-pass ladder — all live, all steerable.
ws → snapshot → backfill → events
steer
Human-in-the-loop, not after-the-fact
Inject a steering message into the agent's next LLM call, pause
mid-pipeline, approve a diff gate, or set a per-run USD budget
that halts the agent the moment spend crosses the line.
pause · resume · stop · message · approve · reject · budget
full
The complete operator surface
Role-to-model lineup strategy, recurring-run schedules, skills,
hooks, agent memory, MCP server config, health diagnostics, cost
observability, and a path-confined file browser — every backend
capability exposed, not just the run lifecycle.
strategy · capabilities · automation · insights · operations