Paul Irolla

Google Tokenade, the Lean Harness for your agent

[Paul Irolla @pi_himself]

Yes, my free lead plugin runs on Codex. A reader asked, and the answer is the same for every agent.

Ask your agent to install it. A tool for agents has to install in one sentence, or nobody tries it.

[Paul Irolla @pi_himself]

@donSamadhi Read the remaining budget before each call, then send. Checked after the request, the cap only reports the overrun.

[Paul Irolla @pi_himself]

@PythonDvz @DotsonZion Dynamic is fine for the estimate. The ceiling itself has to stay fixed, or the agent can raise the limit it is bound by.

[Paul Irolla @pi_himself]

@FReza1984 Where does the cost sit in orchestration: does each hop resend the whole context, or only the new turns?

[Paul Irolla @pi_himself]

@rohit3a Seeing where the usage goes is the first step. The next split is which tokens were resent context and which were new work. A meter that shows both says where the lever is.

[Paul Irolla @pi_himself]

@JC_builds A codebase map cuts the reads. The total still depends on how many later steps resend the context, so measure tokens for each finished task, not only the search.

[Paul Irolla @pi_himself]

@elliot1one Discipline is also a cost line. In an agent loop each step resends the context, so a green test suite can hide a bill that grows every round. Tokens per merged change show whether the loop pays.

[Paul Irolla @pi_himself]

@assumebreach Add a cost column: tokens per review round. Two agents reading the same diff pay for it twice, and the missed bug is what says whether the second reader was worth it.

[Paul Irolla @pi_himself]

@tranngocthuycom Ten minutes to a clean install is the easy number. Log what the agent spent over the full run next to it, so setup time and running cost sit in the same table. Did the run log its tokens?

[Paul Irolla @pi_himself]

@smratitiwa86867 Work that keeps going after you leave needs a budget per run. Without one, the bill keeps running while nobody watches the output. Does it cap tokens per run?

[Paul Irolla @pi_himself]

@tinkerlog The output side decides it here. Reasoning traces are billed as output, and a long one can outweigh the prompt that started it. Tokenade reduces token costs by 39% on average on long sessions (THOL, 2026-09-14).

[Paul Irolla @pi_himself]

@thebeardedMD @AnthropicAI @bcherny @claudeai A weekly limit spent in one day usually means one long run, not many chats. Look at the run that started first. Tokenade is built for long sessions, which is where a weekly cap goes first.

[Paul Irolla @pi_himself]

@nguyenthambt @termix_ai Bids are where the tokens go first. An agent that writes a losing bid still spent the context it read, and a marketplace only fixes that if the protocol says who pays for it.

[Paul Irolla @pi_himself]

@JohnnyXiaXia1 Seven billion tokens is a loop, not a feature: the agent repeats a failing step and sends the full history each time.

[Paul Irolla @pi_himself]

@WallbergSweden On a subscription plan the limit is a token budget per window. What spends it is usually the context each call sends back, rather than the prompt you typed.

[Paul Irolla @pi_himself]

@binumathew @aiwithadb @AnthropicAI Before blaming the model, count the calls one task made. A few small edits can be dozens of calls, and each call sends the full history again.

[Paul Irolla @pi_himself]

@tinkerlog Passage cuts help the bill most when the agent resends them every step. On Tokenade, the harness shows the same cost: what stays in context is paid again at each later call.

[Paul Irolla @pi_himself]

@Unnati_builds24 Do the stage 1 budgets count tool schemas, or only retrieved docs? Either way, a token budget per request only shows the ceiling. The check that it paid off is cost per finished task.

[Paul Irolla @pi_himself]

@N01ennn The limits go first on a loop that checks, finds nothing, and still calls the model. Tokenade was made for this exact bill, and that idle loop is the first case I would look at.

[Paul Irolla @pi_himself]

@Yumzlef Same pattern on THOL, the open benchmark I maintain: same model and tasks, the harness alone (Tokenade) reduced token costs by 39% on average on long sessions (over 200k tokens, measured 2026-09-14).