MEGA · Meta Evaluation-Grounded Adaptation

Optimize your Agent’s
System Prompts.

Autonomous AgentOpt

Start optimizing See the Demo

Why this paradigm is inevitable

Evaluation is becoming non‑negotiable for LLM systems.

Testing was non‑negotiable for traditional software.

Write

Run

Debug

Classic Dev Loop

Write code → Run → Debug

Linear. Deterministic. Green tests means ship

EDD loop: evaluation, run, measure, optimize, repeat

LLM Dev Loop

Define eval → Run system → Measure → Optimize → Repeat

Statistical. Behavior under distributions. Today’s system is not yesterday’s system, and that is the future.

The Solution

See the loop, live.

MEGA Code is a single binary that runs on your machine and opens in your browser. Drop a project folder and it drafts the spec, builds the eval harness, and runs the optimization loop end-to-end. Scroll to step through the real product surfaces.

Start optimizing See MEGA Optimus

Projects

+ New project4 sessions

Active runs

2of 4 sessions

Wins today

7iters ≥ +1%

Tokens

412k−18% vs avg

Best lift

+9.1rag-pipeline-eval

rag-pipeline-evalrunning

iter 7 · p_at_5: 32.1% → 41.2%

~/code/customer-search·3m ago

retrieval-v2complete

iter 12 · ndcg_at_5: 41.2% → 58.7%

~/code/retrieval-v2·14h ago

agent-routingstopped

iter 4 · accuracy: 67.0% → 71.5%

~/code/router-bench·2d ago

tool-selectorrunning

iter 2 · warming up…

~/code/tool-selector·just now

Live activityrag-pipeline-evaliter 7 · +2.3 p_at_5· 3m agotool-selectoriter 2 started· 12m agoretrieval-v2iter 12 · best 58.7%· 14h ago

Project

.mega

Git

prd.md

scan-result.json

eval-rubric.jsonl

feedback/

feedback_iteration_5.json

feedback_iteration_6.json

feedback_iteration_7.json

strategies/

cheatmap_v3.md

mega-optimize/iter 7 · scoring eval setopus-4-7 · 1m 12s · 41k tok

Assistant

Iteration 7 strategy applied: tighten reranker prompt, raise k from 5 to 8.

Running 200-prompt eval set against the new pipeline…

Tool

✓ eval.run · 200 prompts · 41s

score: 0.412 (+0.023 vs baseline) · p_at_5 primary

Assistant

New best. Committing iter 7 and proposing iter 8 strategy…

AskUserQuestion

Approve strategy for iter 8?

Add hard-negative mining to the reranker training set.

recommendedapproveeditskip

iter 7·41.2%+2.3p_at_5▾

Wisdom

Tighten reranker prompt

Raise k 5→8

Hard-negative mining

Cross-encoder rerank

Query rewriting

From 248 patterns

How optimization works

One loop, every run smarter than the last.

Optimization is not a one-shot improvement — it's a continuous cycle that runs until performance stops improving. Each iteration compounds on the last; the system gets better at the job every time it runs the job.

Epoch Start01

Sample seed set

A stable subset becomes the target every iteration has to beat.

Iter 002

Baseline measurement

Score the current pipeline. Every future iteration is judged against this number.

Iter 103

Wisdom curation

The graph assembles a curated orchestration — not a retrieval dump.

Iter 1 → N04

Iterative refinement

Execute, validate, refine. Each pass compounds on the fixed subset.

Epoch Boundary05

Validation test

Confirm gains generalize to unseen data before closing the epoch.

Loop Continues06

New seed sampled

The Wisdom Graph carries everything learned into the next epoch.

See how the loop runs See MEGA OptimusIncludes use cases and a capability comparison against DSPy, GEPA, Maxim, ReLAI.

Introducing · MEGA Tron

One skill layer
for every coding agent.

MEGA Tron unifies, curates, and grades — so every skill call your coding agent makes is sharper than the last.

See MEGA Tron View on GitHub

Roadmap

From workflows to a workflow language.

MDA — Master-agent DSL — sits between natural-language skills and runtime, with a marketplace and optimization layer around it.

Coming Soon

Jun 2026

MDA SDK — Open Source Release

A workflow DSL between natural-language skills and runtime code.

Reasoning efficiency

Structured workflows over freeform prose.

Token efficiency

Operational meaning in a fraction of the tokens.

Bidirectional

Skill.md ↔ MDA ↔ runtime, both directions.

Jun 2026

Agent Marketplace Launch

Reusable executable workflows, observable from the first run.

Portable

MDA moves between projects — no rewrites.

Observable

Execution graphs with traces and outcomes.

Analytics-ready

Logs unlock observability and debugging.

Jul 2026

MEGA on MDA

Master agents run on optimized MDA, not raw codebases.

Token efficiency at scale

Master-agent context shrinks, fidelity stays.

Compositional reuse

Optimizations carry across projects.

Loggable evaluation

Every execution traces — the optimizer measures directly.

Tech Report

From vision to proof.

MEGA — Self-Evolving Agent Optimization Infrastructure.
We show how evaluation evolves into automated optimization.

SkillsBench Pass Rate

46.5%+4.8 vs SkillNet

84 tasks · 11 domains · Gemini 3 Flash

Curation Latency

11.8s−3.2× SkillNet

No LLM calls during retrieval

Aggregate · GPT-4.1 Mini

75.02+5.50 vs GEPA

HotpotQA · IFBench · HoVer · PUPA

Wisdom Graph Pool

4,207PCR-decomposed

Skills · strategies · curation patterns · trajectories

See benchmarks Read the report

Evaluation-driven development is step one.

Automated optimization is the end state.