Date: 2026-09-06 Status: empirical candidate evidence, not architecture authority.

Eliminate manual ChatGPT-window scaling by finding the strongest real independent-agent substrate that Project Brain can control automatically under $0 out-of-pocket spend.

• Real unattended ChatGPT executions: YES.
• Current Plus active-task limit verified from current OpenAI Help: 5.
• Recurring scheduled task maximum frequency: once per hour.
• Current project action: old E13/V208 watches consolidated; native pool now saturated at 5/5 active tasks.
• Active capacity roles now include Brain Continuation, Brain Saturator, Brain Falsifier, Brain Explorer, and combined Bits Frontier Watch.
• Strength: native, already integrated with Brain instructions.
• Limitation: cannot dynamically create arbitrary extra chat windows; unattended tasks can pause; 5 active tasks is hard account limit on Plus.

• Authenticated account: YES.
• Custom team agents created successfully: Spawn Surface Hunter, Agent Factory Architect, Capacity Design Falsifier.
• Personal minimal runtime probe created successfully.
• Actual agent chat execution: FAILED for multiple independent custom agents including personal probe.
• Classification: PROVISIONABLE_BUT_NOT_EXECUTABLE_CURRENTLY.
• Do not count created records as agent capacity.
• Public pricing says Free plan exists with small LLM-message allowance, but current user's actual run path is not functioning, so economics are irrelevant until execution works.

• Managed-agent creation/orchestration tools are available.
• Task execution operations are explicitly marked billable by connector contract.
• Classification: REAL_AGENT_SUBSTRATE_BUT_NOT_DEFAULT_UNDER_$0_OUT_OF_POCKET.
• May be reconsidered only if independently verified prepaid/free credit can be hard-capped to zero additional spend.

• Replit Agent available for creating/updating/inspecting Replit Apps.
• Not exposed as a generic arbitrary parallel mission-agent runtime through current connector.
• Classification: SPECIALIST_APP_BUILDER, not general Brain capacity pool.

• Connected-tool route exists for terminal automation.
• Current empirical result: no authorized device connected at test time, so local Codex supervisor cannot be installed or verified now.
• Classification: POTENTIALLY_STRONG_BUT_HOST_CURRENTLY_OFFLINE.

Current OpenAI product evidence:

• Codex app supports managing multiple agents in parallel.
• Codex Automations can run recurring background workflows.
• Plus includes Codex usage subject to plan rate limits; buying resets exists but project must not use paid resets.
• Strongest candidate route: Brain-controlled Codex agent factory using existing ChatGPT/Codex entitlement rather than API billing.

Unproven implementation questions:

1. Can Codex CLI/app be driven non-interactively under existing ChatGPT subscription on an authorized host?
2. Can multiple independent Codex runs execute concurrently without API-key billing?
3. Can Brain task identity/lease/result evidence be injected/recovered deterministically per run?
4. Can Codex Automations provide persistent re-entry without manual UI interaction, and does the host need to stay online?
5. What exact rate/concurrency limits apply to this user's Plus account?

1. CODEX AGENT FACTORY — strongest candidate, needs live host/automation proof.
2. CHATGPT SCHEDULED POOL — proven fallback, now saturated 5/5.
3. CAMBER — provisionable but run path currently failed.
4. BRAINBASE — real but billable execution, disallowed by default.
5. REPLIT AGENT — useful specialist, not general pool.

No substrate counts as capacity until a real independent run starts, performs a decision-relevant task, returns verifiable evidence, and can be re-entered/replaced without manual task reconstruction.

Brain capacity controller -> acquire every legitimate zero-spend execution slot -> assign unique Brain agent identity -> claim distinct work -> run -> append evidence -> commit -> replace/re-enter on death -> keep capacity saturated.