Launching today

Evidra — The research harness

Launching today

Give your agents a problem worth staying with.

0 followers

Give your agents a problem worth staying with.

0 followers

Visit website

Evidra gives research agents durable memory, parallel exploration, and experimental rigor. Independent agent lanes investigate competing ideas and share discoveries. Failed hypotheses inform future attempts, evidence stays linked to its sources, and replication gates check promising results. Adaptive recovery changes approach after failures, while checkpoints let campaigns resume across sessions. One harness for open-ended research and measurable challenges.

Evidra — The research harness gallery image

Free

Launch tags:Developer Tools•Artificial Intelligence•GitHub

Launch Team

OpenAI

GPT-6 Astra Challenge

How did Astra change the scope or ambition of what you built?

Maker

Evidra began as a CLI for research and coding. Working with Astra helped expand the ambition toward a general-purpose research harness: a system that could investigate a problem, run experiments, preserve evidence, learn from failed approaches, and resume work across sessions. The shift was from helping with individual tasks to coordinating an entire research process. That meant thinking beyond prompts to durable memory, parallel exploration, adaptive experiment selection, independent criticism, replication checks, and recovery when execution fails. It also raised the bar for how we evaluate Evidra itself: proposed harness improvements should be tested across different tasks, with measured decisions about what to keep or revert. Evidra is still an unfinished alpha. The ambition is broader now, but so is the need to demonstrate that these mechanisms improve real research outcomes.

Report

4h ago

Maker

📌

Hey Product Hunt! 👋 Research rarely follows a straight line. Ideas fail, evidence conflicts, and promising results don’t always survive a second experiment. Evidra is a research harness built to manage that process. Some of the techniques we’re bringing together: - Evolutionary search: maintain diverse candidate approaches, select promising ones, and exchange ideas between independent search groups. - Multi-agent exploration: research lanes investigate different formulations, share discoveries, and challenge each other’s conclusions. - Falsification memory: preserve failed hypotheses and negative evidence so future attempts can change course. - Adaptive allocation: use uncertainty, information gain, diversity, and remaining budget to decide what deserves the next experiment. - Evidence provenance: connect source excerpts, claims, hypotheses, runs, and decisions in a traceable research graph. - Critics and replication: review promising findings independently and use repeated experiments and statistical checks before accepting improvements. - Replay and harness evolution: compare alternative search policies using recorded trajectories, and evaluate changes to the harness itself through measured experiments. - Durable execution: checkpoint campaigns across sessions and use failure-specific recovery when experiments hit timeouts, crashes, or resource limits. The goal is one general-purpose harness for open-ended research and measurable challenges, with code execution, local or remote compute, and explicit budgets and permissions. Evidra is currently in alpha under active development. These mechanisms are at different stages of maturity; we’re still validating how well they work together on real problems. We’re not claiming proven superiority over other harnesses. We’d love feedback, contributions, and difficult problems to test against. Which capability would make the biggest difference to your work?

Report

1d ago