Frequently Asked Questions About Lewis Jackson Self-Improving Trading Agent Framework
22 answers covering everything from basics to advanced usage.
// Basics
What exactly makes an agent 'self-improving' versus just automated?
An automated agent executes fixed rules — it will run the same logic on trade one and trade one thousand. A self-improving agent closes a feedback loop: it observes each outcome, scores it toward-goal or away-from-goal, forms a hypothesis, changes one variable, and promotes the winning version to the new baseline. The distinguishing factor is that the strategy itself evolves based on results, not just execution being hands-off.
What is a Sharpe score and why is it used as a goal threshold?
A Sharpe score is a numerical metric representing risk-adjusted profitability — return relative to volatility. The framework uses it as a quantitative success or failure threshold instead of subjective performance judgments. Setting a Sharpe floor (e.g. ≥ 1.2) gives Hermes an objective target to orient improvement toward, rather than relying on 'it feels like it's doing well,' which corrupts the learning signal.
What is a Hermes-readable ledger?
A Hermes-readable ledger is a structured file of all your historical trades — wins and losses — converted into a format Hermes can parse, score, and learn from. It's generated automatically during Phase 3 scaffolding from your existing trade history. This ledger becomes the raw material Hermes analyzes each cycle to identify patterns like oversized losing positions or weak entry signals.
What are score weights and who controls them?
Score weights are configurable parameters in the strategy document that determine how different trade outcomes are scored relative to your defined goal. Hermes owns and adjusts these weights as part of its self-improvement process. They let the agent prioritize what matters — for example weighting drawdown control more heavily than raw return if your failure definition is drawdown-based.
// How To
How do I define a good failure definition?
Make it specific and measurable, not vague. 'Losing money' is not a failure definition; 'drawdown exceeding 15% in any 30-day window' or 'return below 5% over two consecutive months' is. The agent uses this threshold as polarity to orient improvement away from failure and toward success. Without a clear failure edge, the self-improvement loop has no direction and wastes cycles.
How do I feed an existing strategy into the agent?
In Phase 2, choose path A: tell the agent the name of your existing strategy file on your machine. It will locate, parse, and extract your goals and parameters — position limits, entry signals, slippage tolerance — automatically. Phase 3 then converts your trade history into a Hermes-readable ledger. Always review and confirm the generated strategy document before proceeding, since it becomes the source of truth.
How do I complete Railway authentication if the login fails inside the session?
If the interactive Railway CLI login fails inside the Claude Code session, open a split terminal, paste the provided login command, complete browser authentication, then return to the original session and type 'done, continuing'. If you don't have a Railway account, create one during this step — it's free up to a significant usage threshold. This ensures 24/7 hosting is in place before deployment.
How do I verify Hermes installed correctly?
Open a new terminal and type 'hermes'. If it launches, installation succeeded. Hermes is installed automatically as part of the oneshot prompt's handoff phase, so you shouldn't need to install anything manually. Once verified, Hermes becomes the self-learning brain that reviews trades weekly, owns portfolio mechanics and score weights, and writes updated strategy iterations after you approve live mode.
How do I approve a strategy promotion to the new baseline?
Monitor cycle outputs using the check-in commands provided in the final configuration summary. When Hermes produces a cycle result that moves toward your success definition, that version becomes the new baseline. Track directional progress: toward-goal outputs are good signals worth promoting; away-from-goal outputs inform the next hypothesis. For the first live transition, you edit the Hermes trading strategy YAML manually.
// Troubleshooting
Why is my agent wasting cycles without improving?
The most common cause is an impossible success goal relative to your starting capital — the agent cannot self-improve toward an unreachable target. For example, $10 starting capital cannot target $1M/month. Sanity-check your success definition against your capital. Other causes include a vague or missing failure definition, which leaves the loop with no polarity to orient improvement direction.
My results changed but I can't tell why — what went wrong?
You likely changed multiple strategy variables between cycles. If you or a secondary agent modified several parameters simultaneously, the learning signal is corrupted and you cannot attribute the result to any single change. Enforce the scientific method strictly: one variable per cycle. If a secondary agent like Cornelius runs, offset its cadence by 3 days to prevent simultaneous conflicting parameter updates.
The agent went offline and missed trades — how do I prevent this?
You almost certainly hosted the agent locally instead of on Railway. A local-only agent goes offline whenever your machine shuts down or drops network, violating the Reliability criterion. Deploy to Railway for 24/7 cloud hosting so execution never depends on your computer being on. The CLI integration also pushes strategy updates automatically without manual redeployment.
My agent is drawing wrong conclusions from news — how do I fix it?
This is a data accuracy problem. Different AI interpretations of the same news article can produce different conclusions, causing decision drift. Establish objective, rules-based interpretation criteria for all non-numerical inputs before deploying. Numerical and API data should also be validated for reliability — inaccurate API connections corrupt every downstream decision. Clean, consistently sourced data is the foundation of every accurate cycle.
// Comparisons
How does this framework compare to backtesting a strategy manually?
Manual backtesting validates a fixed strategy against historical data once, then leaves you to interpret and adjust by hand. This framework runs live and continuously, letting Hermes observe real outcomes, score them against a defined goal, and iterate one variable at a time on a weekly cadence. Backtesting tells you if a static idea worked historically; this system keeps improving the idea forward without you rewriting it.
How does Hermes differ from just re-prompting ChatGPT for trade ideas?
Re-prompting a general chatbot gives you disconnected suggestions with no memory of prior outcomes, no scoring against your goal, and no controlled experimentation. Hermes natively learns from every engagement, maintains a structured trade ledger, changes exactly one variable per cycle, and promotes winning versions to a persistent baseline. It's a disciplined scientific loop with continuity, not a stream of one-off opinions.
How does the oneshot prompt approach compare to configuring the stack manually?
Manual configuration means separately setting up environment detection, strategy onboarding, scaffolding, cloud deployment, and Hermes installation — each a chance for error and drift. The oneshot prompt collapses all of that into a single copy-paste that Claude Code orchestrates through guided phases, reducing setup friction to near-zero and ensuring reproducibility. It also improves over time as community feedback is incorporated, so newer versions supersede older ones.
How does one-variable-at-a-time testing compare to optimizing everything at once?
Optimizing many parameters at once may improve results faster in a single jump, but you lose attribution — you can't know which change caused the gain, so it doesn't compound reliably. Single-variable testing is slower per cycle but produces a clean learning signal: every improvement is traceable and becomes a durable baseline. Over many cycles, the clean signal compounds into more robust, explainable strategy gains.
// Advanced
Can I run more than one self-improving agent on different assets?
Yes, but manage cadence conflicts carefully. If a secondary agent like Cornelius operates alongside Hermes, offset their review cadences — the framework defaults Hermes to a weekly cycle with a 3-day offset from any secondary agent to prevent simultaneous conflicting parameter updates. Each agent should have its own well-defined goal, ledger, and single-variable discipline so learning signals stay clean per asset.
How does the agent handle the transition from read-only to live write mode?
The first Hermes cycle is strictly read-only — it observes, scores, and produces a markdown review but writes nothing to the live strategy. You then review that output to confirm Hermes understood your strategy and goals correctly. Only after review do you manually flip to live mode by editing the Hermes trading strategy YAML. This gating step protects live capital from a misconfigured loop.
What should be in the four criteria I define before writing any code?
Answer in writing: (1) how you'll ensure data accuracy, (2) how you'll guarantee 24/7 reliability, (3) your exact success definition — target return, Sharpe floor, drawdown limit, and (4) your exact failure definition. These four answers become the scoring and improvement backbone of the entire agent. Skipping them leaves the self-improvement loop with no direction, which is why roughly 90% of trading agents fly blind.
How does the agent sanity-check whether my goal is achievable?
It cross-references your success definition against your starting capital and flags mathematically impossible targets — for instance, targeting 10x returns in 30 days on minimal capital, or $1M/month on $10. If flagged, revise your goal to realistic bounds. This matters because the loop orients improvement toward the target; an unreachable target means every cycle chases something impossible and produces no useful learning.
What happens to my strategy document once Hermes takes over?
The strategy document — covering asset, position limits, slippage tolerance, gas reserve, score weights, target return, Sharpe floor, max drawdown, and failure thresholds — becomes the source of truth the self-improvement loop references. Hermes owns portfolio mechanics and score weights and writes updated iterations each cycle. When a cycle moves toward goal, that revised document becomes the new baseline, so the source of truth evolves with proven improvements.