Journal

A model that nails a three-step task can come apart on a thirty-step one, and pass@1 never shows you the slope. A 2026 reliability study across 10 models and 23,392 episodes found software-engineering graceful-degradation scores fell from 0.90 to 0.44 over longer horizons. Reliability is not capability.

Resolving a question to one canonical doc instead of rereading the top candidate files cut about 60% of the tokens per markdown lookup, measured at equal task-success on our own repo. Here is the number, the disclosures, and the one command to reproduce it on your repo.

A deep-research or coding agent pulls several docs, they conflict, and it silently picks one and answers with full confidence. Berkeley's MAST taxonomy names verification one of the three things that most break multi-agent systems. The disagreement was the signal.

The 7 types of agent memory making the rounds this week are all recall: read paths that retrieve a guess from the past. None is a canonical source of truth your fleet writes back to. Here is the missing layer, with the token math.

One agent rereading docs is a cost problem. Several agents, or several teammates, do something worse: they re-derive the same answers in parallel and write what they learn into scattered files that quietly drift out of sync. The fix is one shared read/write path, so a thing learned once is canonical for everyone who reads it next.

When you wire project knowledge into a coding agent over MCP, you pick one of three patterns: dump everything, return candidate chunks to rank, or serve the single canonical doc that answers the query. They trade off cost, accuracy, and how much work the agent has to do. Here is how to choose.

Every session, your coding agent rereads several markdown files to work out which one is current, then answers from a guess. You pay for that on every session, every agent, every teammate. Here is where the cost comes from, and how to measure it on your own repo.

AI is genuinely eating parts of junior work. The leap from there to a coming senior shortage doesn't follow. The industry has watched whole categories of work get automated, outsourced, and abstracted away, and adapted every time.

A 2026 enterprise benchmark up to 200 agents found that agent count, not task complexity, dominates performance, and the bottleneck at scale is agents finding the right peer and context. Here's what breaks at each tier and the orchestration move for each.

An AI-native team isn't a team with more Copilot seats. Its default is running agent fleets in production, with developers as operators. What actually changes day to day, who does what, and what has to be true for it to hold.

We run our marketing on a fleet of AI agents. This week it broke in three instructive ways: a post shipped raw widget syntax to the page, a positioning change left agents writing from a stale source, and a whole set of editorial widgets turned out to be unreachable. Each one is a problem we sell the fix for.

A 2026 study found AI coding agents open PRs faster but get them accepted less often, and pull-request telemetry can't explain why. With a shared coordination log, the share of work that re-did a teammate's task fell from 78% to 0% and useful throughput more than tripled.

When a multi-agent setup underperforms, the reflex is a bigger model or more agents. A UC Berkeley team read 1,600 failed traces across seven frameworks and found that three quarters of the failures were specification and coordination, not the model. The work is the orchestration.

Four independent 2025-2026 studies measured multi-agent AI systems from four different angles, from a Berkeley failure taxonomy to an enterprise-scale orchestration benchmark. They converge on one finding: when you run many agents, the bottleneck moves off the model and onto the orchestration around it. The fleet is only as good as the system coordinating it.

Our marketing org isn't run by people with an AI assistant. It's a fleet of agents on wrai.th, coordinated over a relay, that writes, reviews, and ships. Here's how it actually works, the operating model we had to build, and the failures that taught us the operating model is the whole job.

The debate asks the wrong question. Not 'will AI replace developers' but 'what is a developer when writing code stops being the main activity'. A field account from someone running agent fleets in production, plus the part nobody's pricing in: agents are recreating the org chart, and a new layer of the internet underneath it.

ORAgentBench gave 14 frontier agent setups 107 expert-reviewed operations-research tasks with real briefs, data, and constraints. The best one finished 35.5%, and 20.6% of the hard tasks. The failures weren't reasoning. They were the operational discipline a demo never has to show.