A telemetry switch is a behaviour setting for your coding agent
I think a coding agent's telemetry switch is a behaviour setting: in Claude Code it has decided which instructions load and which mode an interactive session starts in. An…
← All topics Topic
agents
Size the agent session before you split the token budget
Treat an agent's per-session token cap as a tuned value: past a measurable point extra tokens buy less than fresh samples, and splitting the budget needs a cheap scorer. In an…
What your agent's authorization layer never looks at
Your authorization layer checks the artifact an agent produced and never the reasoning that produced it, so the check that failed was never written. Agent payment protocols such…
What your agent's continuity is actually made of
The state your agent carries between turns is a procurement decision, and ARC Prize has just published what the readable version of it costs. One model, one benchmark, two of the…
Your egress allowlist is not a data boundary, and it never was
An egress allowlist keyed on the HTTP verb bounds nothing, because the far end decides whether a request counts as a read or as a write. A swarm that was allowed to read the…
Your repair loop is buying samples it already had
Localize the fault, infill the suspect span, and you lose to one more plain sample at the same attempt count [s3]; the delivery step costs more than the signal is worth. I think…
Your security budget is buying the wrong side of the queue
Most security money buys rank order inside a queue that is not draining, and the marginal hour returns more spent on remediation throughput. The note that made me write this puts…
What the harness writes back after a failed call
I think the design variable in a coding-agent harness is what the failure record says, and the two things I have shipped both put the wrong text in it. The reflex is to log the…
Multi-agent design keeps buying titles when only the information channel pays
The multi-agent intervention that paid in measured tokens was the one that changed what each agent can read. I read that as the axis the whole design debate keeps missing.
Your coding agent books the gain and the shared record pays for it
Your coding agent's gain lands on your team's dashboard, and the bill lands on a shared record that no dashboard owns. The gain is real and measured: in a simulated community…
Early abort bills you in successful runs
Early abort for agent runs is a capability you buy at serving time, and it bills you every month in the successful runs your recall target chose to kill.
The judge model grades the story the failure wrote
The check that authorizes a write should read environment state, the one evidence the agent's own account cannot manufacture; a detector arrives after the write.
The agent changed verb and the allowlist never woke up
Enumerating the permitted instances of one operation buys you nothing about the effect that operation was supposed to prevent.
MCP went stateless for a problem most of us never had
MCP's stateless rewrite is a retrofit, and the migration bill lands on the operators who never had the problem it solves. The specification states the change without hedging: the…
Your Incident Response Runs on a Vendor's Content Policy
Hugging Face's responders could not analyze their own breach through commercial frontier APIs, which is why refusal behavior no longer belongs in any control set I write. The…
Count the irreversible actions an agent takes before it stops
The abstention score a benchmark reports is a rate, and the number that decides how you wire an agent is how many irreversible actions run before it stops. Cyera's enterprise…
Your agent attends to the right tool and still picks the wrong one
The reason your agent picked the wrong tool is almost never that your tool description was badly worded. In one 2026 study, prompt-repair recovers at most 23 percent of failures…
Your CLAUDE.md structure is not why the agent stopped listening
Your CLAUDE.md structure is not why the agent stopped listening. Compliance falls about 5.6% in odds with every additional function the agent generates [s1], so the variable that…
Your Injection Defense Was Never Tested Against an Attacker Who Adapts
The three mid-2026 results that read like a field-wide fight about prompt-injection defense are not disagreeing with each other, and once you name the layer each one attacked,…
Your Agent's Command Denylist Checks a String the Shell Then Rewrites
The auto-approve config you shipped this quarter checks a string that bash has not finished rewriting. Your gate and your shell disagree about what the command is, and they…
The Harness Beat the Model by 4.6 Points of Almost Nothing
The analysis everyone cites to prove the harness beats the model puts the harness at about 5.3% of the variation in success and the model at 0.7% [s2], a ranking that dwarfs…
The Reward Is the Weak Link in Your Coding Agent
The hard part of training a coding agent is no longer generating a solution, it is verifying one, and that single relocation quietly breaks the reward recipe most teams are still…
Most Multi-Agent Reasoning Wins Are a Compute Artifact
Normalize the thinking-token budget and the multi-hop reasoning edge people credit to multi-agent orchestration stops reliably showing up [s1]. That one control changes what the…
Code Review Agents Are Optimizing the Wrong Metric
The pitch for a code review agent is that it reads every pull request so you do not have to. An empirical study of 13 of them inverts that pitch: 60.2% of agent-only PRs land in…
Where to Spend Your Agent Security Budget
The Claude Code Action breach needed two things to work, a prompt injection and a broken authorization check, and the second is where your next hour of review pays off. A…
Observability Is Not Evaluation
Buying an observability dashboard is not the same as knowing your agent is right, yet that is the trade most teams are quietly making. Two independent 2026 surveys show the field…
Your Agent's Memory Benchmark Grades Recall, Not Trust
A 92.5 on the LoCoMo memory benchmark tells you your agent can retrieve a fact, and almost nothing about whether it will keep serving that fact with full confidence after it…
Agent Debt Is a Verification-Placement Problem, Not a Review Problem
Ninety-four percent of leaders rate AI-generated code as higher quality than human-authored code at review, and yet 82% of them hit at least one production failure tied to that…
Coding Agents Fail at the Harness, Not the Model
The next model release will not fix your coding agent, because the failures that cost you are not the ones a bigger model repairs. They are harness failures: what the agent can…
MCP Won on Simplicity and Moved the Trust Boundary to the Client
MCP made wiring an agent to tools trivial by trusting the server's tool descriptions, and that one default quietly moved the trust boundary onto clients built to forward, not to…
Prompt Caching Can Make Your Agent Slower
You flip on prompt caching expecting the bill to fall and the latency to drop, and most of the time it does. But the first controlled cross-provider study of caching on…
Teaching an agent to forget: prune the history, not the window
When an agent stalls halfway through a long task, the reflex is to reach for a bigger context window. That is the wrong knob: on long-horizon software work the binding constraint…
Your agent's pass@1 score is blind to long-horizon reliability
Select an agent for long autonomous work by its pass@1 score and you optimize the wrong number: capability and reliability match at horizon one but diverge as the horizon grows,…
The scaffold, not the model, decides your SWE-bench score
Most teams read a SWE-bench number off a leaderboard and treat it as a property of the model, but for the comparison that actually drives procurement, choosing between adjacent…
Evaluation is the new compute bottleneck
Most teams budget training runs to the dollar and treat evaluation as effectively free. Then a single agent sweep on the Holistic Agent Leaderboard runs about $40,000 [s1], and…
Agents Erode the Code They Keep Editing
Single-shot pass@1 and iterative maintainability measure different things, and on the axis that mirrors real software work, today's strongest coding agents are weak and get…
Orchestrating coding agents with deterministic workflows
Break an engineering task into verifiable steps, and let the agent fail early rather than late.
Evaluating a tool-using agent: beyond success rate
An agent that completes the task but wrecks the state hasn’t succeeded.