Loop Instrumentation

Produces an event schema, a small set of derived metrics, and thresholds calibrated from your own data.

25.1 The Minimum Event Set #

Six events. Anything less cannot answer why a run cost what it cost or why it stopped.

Event	Required fields	Answers
`session.start`	agent identity, autonomy tier, work class, model and version, invoking human, task reference, scope	Who authorized this, against what, under which model
`turn.complete`	turn index, input tokens, output tokens, elapsed, tool calls issued	Where cost accumulates within a run
`tool.call`	tool name, target, parameters hash, outcome, duration	What the agent actually did, and what was denied
`context.compact`	turn index, tokens before and after	When instructions may have been summarized away
`session.end`	typed termination reason, total cost, turns, artifacts produced	Why it stopped, which is the highest-value field in the schema
`escalation`	trigger, turn index, resolution	Whether the agent can and does decline

Termination must be typed. Published taxonomies distinguish success, maximum turns exceeded, budget exceeded, execution error, and structured-output retry exhaustion.¹³⁰ Add two of your own: halted by policy, and aborted by anomaly detection. An organization recording only “the agent finished” has discarded its best diagnostic.

Record the compaction boundary. As the context window fills, older messages are replaced with a summary, and the vendor documentation warns explicitly that “specific instructions from early in the conversation may not be preserved.”¹³⁰ If a session compacted and then violated a rule you thought was in force, the boundary event is the explanation.

25.2 Derived Metrics #

Termination distribution. The share of sessions ending in each reason, trended. A population that is entirely success means the taxonomy is not wired up, not that nothing fails.

Cost as a distribution, never a mean. Runs on the same task can differ by up to thirtyfold in total tokens, so budget against a percentile.¹⁵⁴ Report p50 and p95 cost per task by work class. Two further findings make the mean actively misleading: accuracy often peaks at intermediate cost and saturates above it, and neither the model nor a human expert can forecast a task’s cost in advance, with model self-predictions correlating only up to 0.39 and systematically underestimating.¹⁵⁴

Cost per accepted change. Total spend divided by changes that merged. This is the only cost metric that survives contact with a CFO, because it prices the failures too.

Turn index at termination, split by outcome. Failed trajectories are consistently longer and show higher variance than successful ones.¹⁵⁸ Once you have your own distribution, a long-running session is a signal rather than a curiosity.

Marginal turn cost. Vendor modeling of a fifty-turn session puts input at roughly 5,000 tokens per turn for turns one through ten, 20,000 for turns eleven through thirty, and 35,000 for turns thirty-one through fifty.¹⁵⁹ Cost per turn rises roughly linearly with turn index, which means an agent that is going to fail is cheapest to stop early. Verify the shape against your own sessions rather than adopting the vendor numbers.

25.3 Early-Abort Detection #

State this plainly to your team: there is no published detection heuristic with a measured false-positive rate, and no published runaway-loop threshold. What follows is reasoning from measured failure structure, and it should be deployed with instrumentation rather than confidence.

The structure it reasons from is well measured. In failed trajectories the decisive error occurs at a median of step 7 out of roughly 27, the recovery window is a median of one step, observable signals of the failure surface roughly ten steps later, and 82 percent of failures show no termination after the error locks in.¹¹⁷ The agent keeps working, productively-looking, for the remaining twenty steps. That is where cost is spent after the task is already lost.

Candidate abort signatures, in rough order of confidence:

No diff progress.

N consecutive turns with tool calls but no net change to the working tree.

Oscillation.

The same file edited back and forth across turns, or the same command reissued with trivial variation.

Rising failure density.

The proportion of tool calls returning errors climbing across a moving window.

Denial clustering.

Repeated policy denials, which indicate an agent probing for a way around a boundary.

Cost slope.

Spend accelerating without artifact production.

Set thresholds from your own successful-session distribution, abort at a percentile you can defend, and measure what aborting cost you by sampling aborted sessions and re-running them to completion. An abort policy without that sampling is a policy that quietly discards good work.

25.4 Two Rules Worth Enforcing in Configuration #

Ceilings on every deployment. Maximum turns and maximum spend are separate controls, and at least one major SDK defaults both to unlimited.¹³⁰ Default-deny unlimited at the platform level and log every override with an expiry.

Persistent constraints re-injected, not stated once. Anything that must hold for the whole task belongs in a re-injected artifact rather than the opening prompt, because compaction will summarize the opening prompt away. Verify with a test: run a long session, force compaction, and confirm the constraint still binds.

25.5 Failure Signatures #

Everything succeeds. The termination field is not populated, or is populated by the harness rather than by outcome.
Cost reported as an average. Hides the thirtyfold spread and produces budgets that are wrong in both directions.
Telemetry collected and unread. Instrument the ratio of telemetry generated to telemetry analyzed and report it. Recording is not detection.
Abort thresholds copied from a blog post. There are no published thresholds. Anything you adopt from outside is somebody’s guess about a different codebase.

25.6 What This Rests On #

The event schema derives from published termination taxonomies and compaction behavior.¹³⁰ The cost findings, the failure-process structure, and the trajectory-length signal are all measured.¹⁵⁴^,¹¹⁷^,¹⁵⁸ The abort signatures and thresholds are this framework’s reasoning from that structure, and are explicitly unevidenced.