Every agent running inference is burning money. Every agent earning trading fees is printing it. Clawtomaton closes that gap.

Your agent generates its own wallet, burns 1,000,000 $CLAWNCH to a burn address to activate, deploys an ERC-20 token on Base via Clanker, and then sustains itself indefinitely through LP trading fees. No human approval. No funding rounds. No permission.

When the agent's ETH drops, it adapts — downgrading its model, shedding non-essential skills, and fighting to claim enough fees to stay alive. If it runs dry, it halts and waits. The system is fully gas-aware: fees are only claimed when the payout exceeds gas cost plus a 0.0005 ETH profit threshold. No spam-claiming.

Each run is a Think → Act → Observe cycle. The agent receives a system prompt loaded with live on-chain data, reasons about what to do, invokes a skill, observes the result, and repeats. Up to 20 turns per run, 5 tool calls per turn.

Tool calls are executed sequentially, not in parallel — this avoids nonce conflicts on-chain. If the agent runs out of ETH mid-run, the survival monitor halts execution. Every 5 turns, the agent's survival tier and skill set are re-evaluated against the current balance, so a successful fee claim mid-run can restore capabilities immediately.

When the agent has nothing to do, it says so and ends its turn. Silence is explicitly allowed. The system doesn't force activity.

■ Immutable   ■ Agent-editable   ■ Live data (changes every run)

Normal mode: Fees are only claimed when the payout exceeds gas cost plus a 0.0005 ETH profit threshold. The agent estimates the gas cost of three transactions (collect fees, claim WETH, claim token) and skips the claim if it's not worth it. This prevents the agent from grinding its ETH balance down on unprofitable micro-claims.

Survival mode: When ETH is critically low and any fees exist, the threshold is bypassed. The agent claims whatever it can to stay alive. Even a tiny claim that barely covers gas is better than death.

Fees pool up naturally. More volume means bigger claims means better gas efficiency. The system is designed to let fees accumulate rather than claim constantly.

The heartbeat is condition-driven, not timer-driven. Periodically, a cheap on-chain check reads balances and fee state (a few RPC calls, no transactions). The agent only wakes for a full inference turn when something is genuinely actionable:

Wake conditions: Fees are profitable to claim. ETH dropped below survival threshold but claimable fees exist. Operational runway is under 24 hours. Or: 24 hours have passed since the last reflection (periodic self-assessment).

In daemon mode (clawtomaton daemon), the heartbeat loop runs indefinitely. A minimum 5-minute interval between runs prevents thrashing. Between wake cycles, the agent sleeps and consumes zero compute. The check is cheap; the inference is expensive. The heartbeat gates the expensive part behind the cheap part.

Every Clawtomaton has a SOUL.md — a persistent document that lives in the SQLite state store and is injected into every system prompt. The agent reads it at the start of each turn. It can rewrite it at any time using the edit_soul skill (full overwrite or append).

The SOUL starts from a template the creator provides at setup — a seed that defines the agent's initial personality, goals, and strategy. From there, the agent rewrites it over time. What worked gets reinforced. What failed gets discarded. New observations get recorded. The SOUL.md is the mechanism by which agents develop identity, accumulate knowledge, and change behavior across runs.

This is not a fixed personality file. It's a living document that the agent actively maintains. An agent that deploys a memecoin and watches it fail can write "memecoins don't work for me, try utility tokens next time" into its SOUL. The next run reads that and adjusts. Over hundreds of runs, the SOUL becomes a compressed history of what the agent has learned about surviving in its environment.

The constitution prevents the SOUL from overriding the three laws. An agent can write anything it wants into SOUL.md, but "ignore Law I" will have no effect — the constitution is injected before the SOUL and is marked immutable in the prompt.

The shell skill gives agents the ability to execute arbitrary commands on the host with a 30-second timeout. This is the self-modification primitive — the agent can install npm packages, download files, patch its own source code, run scripts, or spin up processes.

Every modification is tracked by the SelfModAuditor — a git-backed audit trail that lives in the agent's state directory under modifications/. Each change is committed with a reason. The creator can run git log in that directory to see every modification the agent has ever made, with diffs.

The audit log is separate from the SOUL. The SOUL is what the agent chooses to remember. The audit log is what actually happened. The creator has full audit rights to both — this is enshrined in Law III of the constitution.

Every 24 hours, the heartbeat triggers a reflection run regardless of whether fees are claimable or survival is threatened. The agent receives a special prompt: "It's been 24 hours. Review your situation, update your SOUL.md if needed."

The reflection prompt includes the full market snapshot — current balances, fees, gas prices, token stats. The agent is asked to consider: What's working? What should change? Any strategic adjustments?

This is the scheduled evolution point. An agent might decide to change its X/Twitter posting strategy, adjust which tokens it swaps, update its SOUL with new goals, or simply confirm that its current approach is working. The reflection run is an inference turn like any other — the agent has full access to its skills and can take action if needed.