Polymorphic Agent harness

Polymorphic · AgentsAgent harness · Self-serve

Self-correcting agents for real work

We design, evaluate, and deploy deep research and planning systems that learn from feedback, recover from failure, and systematically prepare your next project. We think this is a robust approach to complex projects.

The plateCreatures: six species of Lenia, a continuous cellular automaton. Each lives by a rule found by searching Lenia's rule space. Gliders travel; still forms drift on a slow current. Wounds: creatures that collide are damaged where they touch. Each repairs using its own rule first and, if it keeps failing, recalls its last healthy state. Click one to wound it.

02 / ComponentsSubstrate + engine
Evidence base
A · Substrate

Search Substrate

We build an evidence base from docs, code, and the web so that every claim is verified and auditable.

  • Citations and provenance tracking
  • Structured design taxonomy
  • Continuous source updates
Search tree
B · Engine

Planning Engine

We use Monte Carlo search over strategies to produce delivery scaffolds and robust cost/timeline estimates.

  • Deterministic sampling of strategy rollouts
  • Scaffolds: interfaces, milestones, tests
  • Sensitivity analysis and risk surfaces

03 / Method

Planning agents that handle failure

We use evidence-grounded rollouts to de-risk scope, cost, and delivery.

Fig. 01

Harness run log

If validation fails, the task goes back to planning with the reason attached. Most failures then ship on a later pass without reaching a person.

Self-correcting
Shipped0Recovered0First pass—To a person0

How to readOne row per task, newest at the top. Each row grows to the right as the task moves through ground, plan, scaffold, and validate. ■ shipped on the first pass · □ recovered · → handed to a person. A simulated run.

  1. Stage 0101

    Ground

    We collect docs, code, data, and constraints, then build a verifiable research base with citations.

  2. Stage 0202

    Plan

    We search strategies using rollouts and sampling, then compare plans against goals and risks.

  3. Stage 0303

    Scaffold

    We generate detailed project scaffolds, including interfaces, milestones, acceptance tests, and owners.

  4. Stage 0404

    Validate

    We backtest projections, run evals, and stress test assumptions, iterating until robust.

04 / AgentsThree in service
Rollouts
P-01

Product Discovery Planner

Our research-driven agent produces evidence-backed briefs, task lists, and risk surfaces for real-world projects.

MCTSResearchCitations
Scaffold
P-02

Architecture Scaffolder

The agent generates production-ready scaffolds: service and data designs, boundaries, schemas, and interfaces.

ScaffoldingEvalsDesign
P50 · P90
P-03

Delivery Estimator

We produce Monte Carlo projections for scope, cost, and timeline. We do sensitivity analysis via probabilistic sampling and scenario rollouts.

MCMCForecastsRisk

05 / Start

Start your project

Tell us about your workflow and what success looks like for you.