The complete archive
55 essays spanning agentic execution, deep reasoning, memory & retrieval, multi-agent systems and the production patterns that hold them together.
55 articles
The Autonomous Dev Framework has hardened into a governed protocol (ADP v1.2): a versioned source knowledge graph, a 7-rule task lifecycle with echo checkpoints, and validation gates that make self-assessed confidence irrelevant.
Read essayContainers, microVMs, WebAssembly: escape resistance versus cold starts versus capability. Ephemeral sandboxes, proxied egress and why sandbox depth is a product feature.
Read essayAmbient authority is why one successful injection inherits everything. Capability tokens, per-task identities and narrowing delegation chains — the 1988 cure, applied.
Read essayDirect is manageable; indirect — the poisoned page, the crafted ticket — is the nightmare. The honest defense stack, and why architecture inherits what attention cannot hold.
Read essayWrite the evaluation first, build the smallest system that passes: red-green-refactor for AI features, from smoke suites to nightly datasets that catch slow drift.
Read essayPosition bias, verbosity bias, self-preference: the documented failure modes of model-graded evaluation — and the calibration stack that makes the proxy trustworthy.
Read essayState change plus event publication in one transaction, relayed at-least-once: deleting the entire category of "the run succeeded but nobody was told."
Read essayBounded queues, value-classed shedding and pauseable generation — propagating "slow down" upstream instead of buffering toward a 3 a.m. OOM.
Read essayOne append-only run log, a dozen purpose-built projections: how command-query segregation lets UIs, auditors and finance teams read the same truth without fighting over its shape.
Read essayClosed, open, half-open — the classic breaker retuned for providers that fail per-model, punish retries with 429s, and bill you for the privilege. With the degradation ladder.
Read essayPolicy lives in a versioned control plane; the agent data plane just works — fast, disposable, and unable to flatter its way past enforcement. Governance as architecture.
Read essayA control plane for autonomous AI coding agents. Give your agent a goal. The loop runs it, scopes it, verifies it, and doesn't stop until it's proven to work.
Read essayDiscover how linear planning and specialized execution cycles are pushing AI towards autonomous, real-world task completion.
Read essayAllocate maximum computational budgets to complex tasks to expand thought horizons and uncover step-by-step conceptual proofs.
Read essayEcho chambers, thrashing, responsibility diffusion, protocol drift: the empirical taxonomy of fleet failures — and why every fix is architectural, not prompting.
Read essayMinsky called it in 1986: intelligence is the organization of simple agents. Modern fleets confirm it — the B-level coordinators, not the workers, set the character.
Read essayWhy recursive self-correction and internal adversarial debate yield highly robust and production-ready intelligent responses.
Read essayRotating blind reviewers, seeded-defect calibration and typed findings: turning "an agent reviewed it" into measured catch rates per defect class.
Read essayTyped handoff contracts, readback acceptance, and authority boundaries — why the interfaces, not the agents, are where multi-agent systems win or die.
Read essayO(n²) chatter, saturated hubs, spanning-tree org charts: how the shape of agent-to-agent communication decides cost, auditability, and the ceiling of the fleet.
Read essayThe four-eyes principle for agents: the one who does the work is never the one who signs it off. Information asymmetry as the design, not a bug.
Read essayContext focus, tool authority and evaluable standards make specialization real; personas without partitioned work are theater. How to tell the difference.
Read essayMajority voting only beats its average member when votes are independent — and correlated LLM voters are not. Quorums, weighted reputation and protected dissent.
Read essayAnnounce, bid, award: the contract-net protocol rebuilt with LLM workers — pricing capability, honesty through reputation, and the anti-thrash rules markets need.
Read essayAdversarial argument under a rubric-bound judge surfaces solo reasoning's blind spots — with honest caveats about eloquence bias and premature concession.
Read essayVersioned input-output fixtures, deterministic assertions, calibrated model graders — the testing discipline that turns prompt changes from vibes into engineering.
Read essayTruncation amputates; compaction curates. Structured state artifacts, layered summaries and pinned cores — curriculum design for the agent you are becoming at step 200.
Read essayPlan files, scratchpads and durable notes: why agents that think on paper outperform agents thinking in-window — and the lifetimes that keep each artifact honest.
Read essayMemGPT mapped OS virtual memory onto tokens: a visible main context, a searchable external store, and self-directed paging. The inversion of control behind long-horizon agents.
Read essayThree generations of retrieval-augmented generation: corrections around a fixed pipeline, modular swappable stages, and retrieval as a tool the model wields strategically.
Read essayLost-in-the-middle degradation, double-billed tokens, and utilization metrics: treating the context window as a budgeted cache instead of infinite memory.
Read essayQuestions and documents are different distributions. Multi-query, step-back prompting and hypothetical-document embeddings fix the mismatch at its origin.
Read essayParent-child retrieval, hierarchical indexes and contextual retrieval — why the unit of meaning you embed decides the ceiling of the whole RAG stack.
Read essayDense misses identifiers, BM25 misses paraphrase, graphs rot without governance. The production funnel: hybrid candidates, reranker judgment, graphs only where relations are the product.
Read essayWorking set, episodic journal, semantic store — and the consolidation process between them that separates memory systems that help from stores that just grow.
Read essayA small model with smart inference-time spend beats a giant at equal FLOPs — if the compute is allocated to the right strategy. The operator's guide to the new frontier.
Read essayAccuracy collapses past a complexity threshold, correct answers get second-guessed into errors. Measured, diagnosed and countermeasured: deliberation as a budget.
Read essayRetrieve structurally similar solved cases, map their relations, adapt what breaks. Structure-mapping theory meets case-based reasoning in LLM systems.
Read essayLeast-to-most, decomposed prompting, plan-and-solve — every accuracy technique is decomposition in disguise, until composition errors bite. How to split well.
Read essayVerification is cheaper, parallel and more reliable than generation — if the verifier is independent, cheap, and hard to game. The economics of checking AI output.
Read essayTrain small models on verified reasoning traces from frontier teachers — and learn what distillation cannot transfer: hesitation, backtracking, and self-correction.
Read essayThinking tokens are a runtime purchase. How budget forcing — from the s1 "Wait" trick to tiered escalation — turns reasoning depth into an operations dial.
Read essayExploit the generator-verifier gap: every answer ships with a checkable artifact — tests, proofs, source spans — and only verified conclusions ship.
Read essaySample k reasoning paths, vote on the answer. The cheapest member of the test-time-compute family — plus the budget curves and vote-normalization caveats.
Read essayExpand candidate thoughts, score them mid-flight, prune dead ends, backtrack on collapse — how inference became search, and when the quadratic cost pays off.
Read essayPersist every step, replay any trajectory, fork to try a different strategy. How durable execution turns unrepeatable AI runs into debuggable ones.
Read essayAgents retry more than any software ever built. Idempotency keys, conditional writes and state-declared actions keep those retries from duplicating reality.
Read essayMulti-step agent actions break atomicity. Sagas — local transactions with compensating inverses — turn catastrophic partial failure into clean unwinding.
Read essayLLMs propose decompositions; classical HTN planners validate and execute them. The hybrid gives agentic planning soundness it cannot get from prompting alone.
Read essayHearsay-II solved opportunistic problem-solving in 1975 with a shared board and independent specialists. Modern agent systems are rebuilding it on purpose.
Read essayGive an agent a fence with full freedom inside and zero authority outside. Why autonomy scales with containment — and how to size the fence honestly.
Read essayDurable queues, event-sourced run state and crash-proof agents: why the in-process loop is the first thing to die in production — and what replaces it.
Read essayCollapsing O(n²) agent chatter into a hub-and-spoke topology — routing, context isolation, typed contracts and the single-point-of-failure trade-off.
Read essaySpend tokens up front on a structured plan, execute it mechanically, and replan only when reality falsifies it. The economics of look-ahead in agent design.
Read essayWhy interleaving written thoughts, tool calls and observations became the load-bearing loop of production agents — and how to make it terminate reliably.
Read essay