deixic
Log inRequest a demo

Blog

AI agent operations

Notes on cost, access, approvals, identity, and the failures that show up after launch.

RSS

Series

Focused collections

Start with the problem you are trying to solve.

Series

AI content spend

Trace model cost, review time, rework, and the assets that never make it to publish.

Series

Agent approvals

Choose which agent actions can finish automatically and which need a human decision.

Series

Identity and access

Keep the owner and permissions visible when agents call tools or hand work to other agents.

Series

Production failures

Find the failures that surface after an agent has already changed something.

Engineering

How we turned our coding agent into our own bugbot

We already had a coding agent that could read a repository, trace callers, and run tools. The reviewer became a second operating mode of it: pinned revisions, forced investigation, an adversary for every high-severity finding, maintainer adjudication, and replay against bugs we already shipped.

· Engineering · 9 min

Engineering

Why we built our own agent sandbox

Daytona, Modal, and E2B can start isolated compute quickly. Our evaluation runtime needed a one-shot lease with tenant and execution binding, plus provider-confirmed destruction.

· Engineering · 5 min

Engineering

After the outage: making agent completion provable

An agent turn can be accepted, scheduled, streamed, and still fail. We changed the runtime so intermediate progress can no longer masquerade as completion.

· Engineering · 6 min

Engineering

A 1-second submit hid a 30-minute outage

How a bounded, read-only agent investigation separated a compound runtime outage into five causal defects, and what we built to make that method repeatable.

· Engineering · 7 min

Cost

Who owns the AI budget when one task crosses three teams?

A practical ownership model for shared AI work: separate service owner, task owner, budget owner, approver, and finance reviewer before costs cross team boundaries.

· Cost · 4 min

Earlier posts

  1. Finance needs an AI spend reconciliation before month-endCost · 4 min
  2. Your agency invoice hides the cost of AI-assisted contentCost · 4 min
  3. Brand review is a production cost, not a final opinionOperations · 5 min
  4. The cheapest content request is the one you do not have to repairCost · 4 min
  5. Your AI budget is wrong before finance sees the invoiceCost · 9 min
  6. What to ask when AI content spend is over budgetCost · 4 min
  7. Every campaign variant has a cost, even when the model is cheapCost · 4 min
  8. Your AI content program is paying people to clean up cheap draftsCost · 9 min
  9. Marketing needs a data boundary before it buys another AI toolIdentity & access · 4 min
  10. Every content exception needs an owner before it becomes a costApprovals & policy · 5 min
  11. AI localization is cheap until every market edits itCost · 4 min
  12. The cheapest model can make your content bill biggerCost · 5 min
  13. Published volume is a weak measure of AI content productivityOperations · 4 min
  14. The AI content vendor checklist finance wishes marketing hadCost · 5 min
  15. The last mile is where AI content gets expensiveOperations · 5 min
  16. Content refreshes can cost more than new pagesAudit & compliance · 4 min
  17. The review queue is where your AI budget goes to dieCost · 9 min
  18. Your AI program has a review queue, even if you call it approvalOperations · 4 min
  19. The rework rate is the metric your AI dashboard is missingCost · 4 min
  20. AI saved us time. Why did the marketing budget grow?Cost · 6 min
  21. AI makes content scope creep hard to seeCost · 5 min
  22. Stale sources turn AI content into expensive cleanupAudit & compliance · 4 min
  23. One API key cannot explain a marketing budgetCost · 7 min
  24. Five AI tools can make one content process more expensiveCost · 4 min
  25. Cost per token is not the cost of a finished assetCost · 5 min
  26. Before you renew an AI content tool, ask what got cheaperCost · 4 min
  27. The content you paid for but never publishedCost · 6 min
  28. Who can use AI for marketing work—and who can approve it?Identity & access · 4 min
  29. The evidence packet marketing needs before publishing AI-assisted workAudit & compliance · 4 min
  30. NIST AI RMF and ISO/IEC 42001 for the agent runtimeAudit & compliance · 5 min
  31. Controls that hold when the agent stops agreeingApprovals & policy · 4 min
  32. Agent spend, before the invoiceCost · 3 min
  33. Discovery lag: how long before you find out what your agents did?Operations · 5 min
  34. What an agent audit row should carryAudit & compliance · 9 min
  35. Why service-account permissions don't fit agentsIdentity & access · 8 min
  36. Five things that keep breaking when AI agents move into productionOperations · 8 min
  37. Policy at action time: what the gate evaluatesApprovals & policy · 10 min
  38. Sub-agent identity: inherit and narrowIdentity & access · 8 min
  39. The lethal trifecta: where it actually livesIdentity & access · 7 min
  40. Stopping a running agentOperations · 8 min
  41. What the EU AI Act means for the agent runtime: an Article-by-article readAudit & compliance · 9 min
  42. Designing approvals operators actually readApprovals & policy · 9 min
  43. Replay: reconstructing an agent action from the audit logAudit & compliance · 9 min