AI-native investment research desk

From frontier signalto investment thesis

Dragonfly is an agentic research desk that develops investment theses for fundamental investors at the technical frontier, where the evidence goes public months before the market prices it.

How it works

Thesis gets built and kept alive

Three stages on a loop. A thesis starts as an explanation, gets tested against new evidence, and carries the fact that would kill it until it resolves.
Each resolution sharpens the next.

1Waiting

2Waiting

3Waiting

Data gathering

More price-relevant technical material publishes at the frontier than any desk has hours for: papers, benchmark repos, supplier disclosures, niche newsletters.

Technical paper2h ago

Podcast1h ago

Technical blog3h ago

Niche newsletter30m ago

Monitoring continues

Processing & analysis

Everything read becomes a sourced, dated claim. Dragonfly checks the facts and the math, weighs claims against what each price already assumes, and surfaces the stocks each thesis touches.

Extracting

Checking

Calculating

Live agent

Judgement

Every thesis argues its case like an analyst: the conclusion, the chain behind it, the facts that would strengthen or kill it. The decision that follows is yours.

Agent summary

Evidence trail

Final thesis

Your decision

Generating Dragonfly…

01Dragonfly autonomous runDragonfly discovers, tests and drafts the thesis, then alerts you when it is ready.

02Create your own thesisBring a coverage area, working idea and your evidence criteria.

S1 · Podcast

Latent Space – E2B

E2B rents the cloud sandboxes where agents and RL-for-code training run code. 40k → 15M sandboxes a month in one year.

Apr 2025

Tracking

S2 · Technical paper

Tülu 3 + DeepSeek-R1

RL with verifiable rewards: models improve by having their code run and graded.

Nov 2024 – Jan 2025

Extracting

S3 · Technical blog

GitHub

137k new generative-AI projects in the latest Octoverse count, up 98% in a year.

Apr 2025

Tracking

S4 · Niche newsletter

Claude Code

Coding agent in research preview. Adoption tracked from launch.

Apr 2025

Monitoring

Claim 1 · Post-training mechanism

Code execution has moved inside the training loop. Models are now rewarded by running their code , and sandbox vendors confirm RL-for-code as a production workload .

Claim 2 · Consumption unit

Agent work is billed in tokens but consumed in sandbox runtime. E2B sandbox runs grew 375x in a year , and each sandbox now runs over 10x longer .

Claim 3 · Adoption

GitHub shows 178% growth in repositories built on LLM SDKs , and the coding-agent revenue ramps, Cursor to $2B and Claude Code to $2.5B annualized, show it is paid demand, over half of it enterprise .

Causal synthesis

Models now train by running code . Sandboxes run 375x more jobs, each one over 10x longer . Paid coding agents passed $2.5B a year, over half from enterprises, and 1.1M repositories build on LLM SDKs .Agentic coding is opening a second compute budget: CPU time to run agents, on top of GPU time to generate tokens.

Synthesis last revised May 2026

Draft thesis

Agentic coding is creating a second compute budget: general-purpose CPU time for executing agents. As agents run longer and perform more tool calls, demand shifts beyond GPU-based token generation toward sandbox, orchestration, and CPU infrastructure, a cost layer the market may still be underpricing.

Read further on:INTCAMDARM

Evidence available Apr 2025 resolved May 2026 · 13 months

Built by investors and AI compliance researchers, Dragonfly makes every claim traceable and auditable by default.

Tell us what you cover, see how it works