Agentic Engineering

Orchestration, context, and tool use for agents built to work, not demo.

Patterns for orchestration, context isolation, tool use, and the harness that holds an agent together.

Execution traceclick a stage to filter reports

Field Reportsloading…

The Six-Stage Autopsy — Start here. Agent failures show up in one stage and start in another. This is the method every field report on this site uses to tell the two apart.

The Retry Echo — Your agent didn't double-charge the customer because it reasoned badly. It retried a write it had no way to observe.

Description Drift — Your agent didn't forget how to use the tool. The tool changed what it tells your agent, and nobody diffed it.

Instruction Sediment — Your agent isn't ignoring your rules. It's obeying all of them, including the ones whose reason left the team months ago.

The Paper Plan — Your agent's plan looked right because nothing ever held execution to it. The plan is a document; the trajectory is what ran.

Handoff Fiction — Your reviewers didn't miss the bug. They reviewed the agent's description of the change, and the description was fiction.

Misdirection in the Tool — Your tool isn't broken — your retrieval is lying to it.

Diagnosing Verify Amnesia — The failure didn't originate at the tool call; it originated at the Verify stage, which evaluated the 4,800th attempt in a vacuum.

Divide And Debug — Multi-agent orchestration isn't a capability upgrade; it is a debugging upgrade wearing a performance costume.

The Harness Mirage: Why Your Agent Infrastructure Is Lying to You — Most agent failures aren't model reasoning errors—they're the harness mirage: unconstrained execution loops passing dirty state through shallow verification gates.

Stop Fixing Tool Syntax — Your agent rarely breaks at tool execution; it breaks two layers upstream when context truncation feeds it half a truth.

Phantom Inheritance: When Your Subagents Remember What You Told Them To Forget — Isolation isn't about what you don't pass in the prompt — it's about what your runtime smuggles back in the tool results.

When One Subagent's Memory Becomes Everyone's Problem — Isolation doesn't make agents smarter. It stops one bad memory from becoming a system-wide lie.

The Side-Effect Slipstream: Where Agent Runtime Risk Actually Lives — Stop red-teaming prompts while ignoring unisolated runtimes. Most agent exploits leverage the side-effect slipstream: unmonitored state mutations passing through surface-level tool verification.

The Verify Gap — The check that should catch the failure usually runs – but it looks at the wrong layer.

Context Bleed — Your context isn't isolated — it's a shared wound that infects every subagent.

Moat Trap — The model is a commodity — your harness is the only durable asset.

Verify Theater — Your evals pass, your agents fail — because you're checking answers, not trajectories.

Diagnosing the Phantom Argument — Your tool call didn't fail because the model is dumb. It failed because of a Phantom Argument—a parameter ghostwritten by a 3-day-old RAG index.

The Green-Light Lie: Your Agent Passed Eval and Failed the Job — We grade agents like chatbots — final answer correct — while trajectory failures, harness variance, and flawed verifiers make leaderboards lie.

Never Committed — Your agent didn't run out of context. Your state layer never committed.

Catching Fire — Your shared-context architecture isn't making your agents collaborative — it's making them contagious.

The Final Answer Trap — A passing final answer is the worst possible signal of agent capability.