Blog — Meharban Singh

Posts

• 2026-09-30My AI Agents Cannot See Their Own Limits. So I Built Them a Window.When agents hit usage limits or context fills mid-task, they stop with no warning. I made the limits visible, and I made the pause independent of the model.
• 2026-09-26My Agent Looked Alive for 36 Minutes. It Had Written 39 Bytes.A stdin-hung worker taught me to measure output movement instead of trusting process existence.
• 2026-09-26A Waived Gate Is a Failed GateA waiver preserves a failure, its reason, and an explicit override trail.
• 2026-09-26Two AI Agents, One Migration ChainDuplicate migration heads taught me to declare disjoint paths and serialize the chain.
• 2026-09-26 The Most Honest Answer My Tool Gives Is “Unknown” A real SessionRx incident: three denominator traps turned context evidence into false certainty, so the rule became unknown rather than a guessed percentage.
• 2026-09-17 Beyond Coding Agents: Why Enterprises Need a Vendor-Agnostic Agentic Software Factory An architecture I'm designing and exploring: a vendor-neutral control plane that governs many coding agents, from many vendors, under one identity, policy and audit trail.
• 2026-09-15 The Topology Rule That Predicts Whether a Stuck System Escapes A research paper, not an incident: why a map of possible transitions predicts escape from a stuck state far better than local energy does — paper included.
• 2026-09-09 How the Factory Learns From Its Mistakes Operator corrections and closed defects become lessons injected into the next prompt. The mechanism, its evidence, and what remains open.
• 2026-09-08 Context Is a Finite Resource: I Built a 24/7 Autonomous Handoff Protocol A handoff protocol for context degradation and zombie sessions: arm, quiesce, save, spawn, adopt, verify, then retire.
• 2026-09-05 Where the Tokens Went A measured account of reducing repeated AI-agent context without compromising verification quality.
• 2026-09-02 When AI “Cleans Up” the Past AI agents can fix a truthfulness problem while quietly removing a useful legacy capability. The controls that make behaviour preservation part of autonomous development.
• 2026-09-01 I Built an AI Software Factory. Here Are the Failure Modes Nobody Puts in the Demo. How a real multi-agent build system fails: stale state, tests that can't turn red, reviewers agreeing without new evidence, and what fixed it.
• 2026-09-01 My AI Agents Passed Every Test. The Software Was Still Wrong. Four failures from building a risk platform with AI agents: a UAT that never committed, a mutation nothing tested, a caller nobody wired, a metric that lied.
• 2026-09-01 'Done' Is a Claim, Not a Fact An installer shipped green with two unsubstituted placeholders. The fix that closed it required a different verifier, a RED/GREEN probe, and a database that refuses self-certified closes.
• 2026-09-01 The Writer Never Approves An orchestrator wrote 2,968 lines directly, then reviewed its own work. A bug in its own checker survived that review. What the role lock that followed actually changed.
• 2026-09-01 The Day I Rotated 15 Credentials A names-only secrets listing printed real values into a transcript, a redaction attempt made it worse, and the rule written afterward was broken again six days later.
• 2026-09-01 Shipping in a Domain I Knew Nothing About Building Astro Rattan meant shipping Vedic astrology without knowing the subject. The fix: take the model out of the answer path and let an expert verify.
• 2026-09-01 I Don't Prompt AI. I Run a Factory. A blueprint reviewed by a second model, hooks that block instead of advise, gates that read evidence not summaries, and a sentinel watching production — measured against 893 tracked issue rows.
• 2026-08-26 An Instruction Is Not a Guarantee Why a spec isn't enough once an agent does real implementation — and the process (skills, hooks, gates, independent review) I built around Claude Code instead.

Most posts take one real incident from the issue ledger — the failure, the root cause, and the rule or hook it became; the occasional fictional one is labeled as such on the page and holds to the same structure.