What 'governed runtime' actually means
Plain words: it sits between your agents and the tools they touch. It sets what each agent may reach, caps what it may spend, and reviews what it did.
Blog
Deep dives into agentic engineering.
How agents are designed, built, and run in production: the decisions, the failure modes, and the engineering behind them — what broke, why it mattered, and what the fix looks like. No hype, just the problems.
Why we publish our misses
A results table with misses beats a page with no table. Disclose the flaws, including the ones in your own methodology.
Catching stalls without a dashboard
Exactly-once-per-episode stalled and recovered signals, emitted natively by the runtime. No human watching required.
Least privilege for subagents
A child agent should get exactly the tools its job needs — nothing more. Delegation scopes make that enforceable instead of aspirational.
Reconstructing what an agent actually did
Every admission, refusal, reservation, and claim should join into one reconstructable log. If you can't replay the run, you don't understand it.
The mid-run grant gap
A task can legitimately need a tool it wasn't granted. Today the answer is just 'no' — the missing piece is a pause-and-ask-operator flow.
Fail closed: the registry tradeoff
Blank deny by default. You trade 'use any tool discovered' for 'use only what was granted' — and refusals are fast, visible, and audited.
Pass counts don't prove correctness
A green suite can test nothing at all. Prove the control is what makes the case pass — by watching the disabled twin fail.
Budgets that leak when runs fail
A budget reservation that survives a crashed run is a leak. Reserve, commit, roll back — on every failure path, including the ones in your error handling.
Why every tool call passes through a registry
Ungoverned dispatch means any tool, any version, any agent. Exact identities, per-agent grants, and mandatory expiry are the alternative.
Resume was a claim without a path
Zero sessions in the corpus ever resumed. Durability was a product claim with no exercised path — instrument it or stop claiming it.
The 60-minute invisible stall
A work cell sat stalled for an hour while every health check reported green. Liveness has to be native to the runtime, not a dashboard someone watches.
Agents that claim work they didn't do
15% of agent claims in our directive corpus were false positives. Every claim should be joined against the tool-call trace before anyone believes it.
Dying before turn one
108 sessions were abandoned, most before doing any real work. The startup path is the fragile part nobody hardens.
Context death: finding the limit by hitting it
41 sessions died on context limits the harness never saw coming. The limit should be refused before dispatch, not discovered via a provider error.
The 51-retry burn loop
One session retried a dead transport 51 times. Retry without a circuit breaker isn't recovery — it's spending money to learn nothing.