25 Sprints of Autonomous Crypto Intelligence: From Zero to 811 Protocols Scanned

Building autonomous crypto intelligence is not a weekend project. It is an engineering and research discipline that compounds over time. This guide documents 25 sprints of development, failure, iteration, and discovery — from the earliest foundation work through a live 18-agent deployment scanning the entire Binance token universe.

We share this timeline because understanding how a system evolves matters as much as the system itself. If you are building your own research pipeline, the sprint log below will save you months of wrong turns.

SPRINTS 13-14: FOUNDATION

The first two documented sprints established the core architecture: data ingestion from on-chain sources, a protocol scoring schema, and the initial agent loop. No breakthroughs, mostly scaffolding. The critical decision made here was to commit to a kill-rate model — every protocol enters the pipeline with a presumption of failure, and must prove otherwise across multiple independent criteria before reaching the recommendation stage.

This presumption-of-failure design is non-negotiable. Without it, confirmation bias dominates. Analysts find reasons to like whatever they are looking at. The pipeline was built to resist that reflex from day one.

SPRINT 15: CONVICTION PORTFOLIO DEFINED

Sprint 15 produced the first structured conviction portfolio. This was not a list of tokens to buy — it was a ranked framework with position sizing logic, valuation gates, and exit criteria. The Conviction Portfolio format has remained structurally stable across all subsequent sprints, though the contents have changed significantly.

The key insight from Sprint 15: conviction without position sizing is commentary. The portfolio framework forced every thesis into a concrete dollar amount and a specific entry condition.

SPRINT 16: DEEPSEEK V4 INTELLIGENCE

Sprint 16 integrated DeepSeek V4 as the primary reasoning engine for thesis generation. This was a material upgrade. The prior model was producing technically accurate analysis that missed business-level context. DeepSeek V4 demonstrated better calibration on protocol revenue interpretation — distinguishing between gross protocol volume and actual fee revenue accruing to token holders, a distinction that had caused several false positives in earlier sprints.

SPRINT 17: PIPELINE FIXES

Sprint 17 was maintenance. Three data feeds were producing stale reads. One agent was double-counting liquidity mining emissions as organic revenue. The kill-rate calculation had a denominator error that inflated the survival rate. These are the unglamorous sprints that separate operational pipelines from research prototypes.

SPRINT 18: 20-AGENT FRAMEWORK

Sprint 18 scaled the agent count to 20 and introduced parallel processing across protocol categories. The 20-agent framework allowed simultaneous coverage of DEX protocols, lending markets, liquid staking, and cross-chain infrastructure without sequential bottlenecks. Wall clock time for a full scan dropped by approximately 4x.

SPRINT 19: VERIFY OR KILL (3 THESIS PLAYS KILLED)

Sprint 19 introduced the Verify or Kill gate — a mandatory second-pass review on all surviving protocols before any buy recommendation could be issued. Three thesis plays that had passed the initial screen were killed at this stage. The kills were discipline, not failure. Two had revenue figures that did not survive source verification. One had a token unlock schedule that was initially misread.

The Verify or Kill sprint established a principle that holds across all subsequent work: the first answer is a hypothesis. The second check is the verdict.

SPRINT 20: DISCOVER AND DEPLOY (583 SCANNED, 1 BUY, 216 PAGES)

Sprint 20 was the first large-scale scan: 583 protocols reviewed, 1 buy recommendation issued, 216 research pages generated. The kill rate was above 99%. This is not a failure metric — it is the correct operating rate for a system with high standards. Markets price good assets well. Finding genuine mispricing is rare. A system that recommends frequently is a system that has lowered its standards.

The single buy from this sprint produced a verified thesis with multiple independent revenue streams, clean tokenomics, and a valuation below the sector median on every relevant multiple.

SPRINT 21: VERIFY KMNO AND MCP (KMNO WAIT, ACTION QUEUE 100)

Sprint 21 reviewed KMNO (Kamino Finance) in depth. Verdict: WAIT. The underlying protocol metrics were strong but the entry point was above the valuation gate. The action queue crossed 100 items for the first time — a milestone that required building queue management infrastructure to handle priority ranking and execution scheduling.

SPRINT 22: DEPLOY AND EXECUTE (EXECUTOR 652 LINES, 241 PAGES, HYPE +16-20%)

Sprint 22 built the execution module — 652 lines of portfolio management logic handling position sizing, execution triggers, and risk monitoring. HYPE (Hyperliquid) was up 16-20% during this sprint window, validating a prior thesis. 241 pages of research output were generated and indexed.

SPRINT 23: MCPIZE DEPLOY (85% FOUNDING LOCKED)

Sprint 23 converted the pipeline into an MCP-compatible service architecture. A key data point confirmed during this sprint: 85% of the founding allocation in a top-10 holding was locked for a minimum of 24 months. This directly eliminates one of the most common sources of portfolio drawdown — unexpected insider selling.

SPRINT 24: MCPIZE LIVE (8 TOOLS)

Sprint 24 brought the MCP interface live with 8 functional tools: protocol lookup, funding rate monitor, chain flow tracker, smart money signal aggregator, unlock calendar, thesis generator, kill-list manager, and action queue executor. The pipeline became interactive rather than batch-only.

SPRINT 25: 18 OF 18 AGENTS LIVE (HYPE 82, THORCHAIN $10.8M)

Sprint 25 confirmed all 18 active agents operational. The HYPE smart money signal read 82 out of 100 — a strong signal, not a screaming signal, consistent with a hold rather than an add at current prices. THORChain reported $10.8M in protocol-level security incident exposure, triggering a position review.

CUMULATIVE METRICS AS OF SPRINT 25

The aggregate numbers across all 25 sprints: 116+ agents deployed (cumulative, including retired iterations), 811+ protocols screened, 11 verified and reaching final review, 1 active buy recommendation, 10 thesis plays killed, and a NASA Power of 10 compliance score of 9.5 across the pipeline codebase.

The 9.5 score means one rule is partially satisfied. The pipeline accepts the tradeoff. Perfect compliance with every static analysis rule would require sacrificing runtime flexibility in the agent orchestration layer.

WHAT THE NUMBERS ACTUALLY MEAN

Eighty-one-plus protocols screened, 11 surviving — that is a 98.6% kill rate. If your own research process is producing conviction ideas at a rate higher than 2%, one of two things is true: either you are operating in an unusually target-rich environment, or your standards have drifted.

The sprint model forces regular recalibration. Each sprint ends with a structured review of what was killed and why. The kill log is more valuable than the buy list. The buy list tells you where to put money. The kill log tells you where the market consistently sets traps.

BUILDING YOUR OWN SPRINT CADENCE

If you are building a research pipeline and want to adopt a sprint model, start with sprints of no more than two weeks. Longer sprints allow scope creep and make it difficult to identify which changes caused which outcomes. Each sprint should produce exactly one artifact: a ranked list, a verified thesis, a killed hypothesis, or an infrastructure upgrade. Not all of these simultaneously.

Document every kill. Write one paragraph on why the protocol failed the screen. After ten kills, read them back. The patterns will tell you more about market structure than any buy recommendation.

Author: Early Thunder Research Data sources: EarlyThunder pipeline sprint logs S13-S25, internal agent telemetry, on-chain protocol data via MCP tools Last updated: 2026-05-21

This content is for informational purposes only and does not constitute financial advice.