Agentic workflows

Agentic workflows sit at the intersection of a technical problem and an organizational one. The technical problem is getting an LLM to take a sequence of actions reliably across many steps without losing state, making unrecoverable mistakes, or silently lying about completion. The organizational problem is that faster code generation does not fix the underlying bottlenecks of shared context, specification clarity, and management coherence — it amplifies whatever alignment or misalignment already exists the-bottleneck-was-never-the-code.

On the engineering side, the recurring lesson across multiple sources is that prompts are not the right lever for reliability. A data engineering agent iterated through three architectures before settling on a single general-purpose agent governed by environmental constraints: tool design, stable ID keys, and context visibility dont-prompt-your-agent-for-reliability-engineer-it. Brian Suh reaches the same conclusion through a different route: complex tasks need deterministic control flow encoded in software, with explicit state transitions and validation checkpoints, not increasingly elaborate prompt chains agents-need-control-flow-not-more-prompts. Anthropic’s harness engineering guide formalizes this into five subsystems: instructions, state, verification, scope, and session lifecycle walkinglabslearn-harness-engineering.

State management is where many agent systems break down. Anthropic’s 12-factor-agents project argues for unifying execution state and business state into a single context-window-derived thread, making the system trivially serializable, debuggable, and resumable from any checkpoint humanlayer12-factor-agents. The practical version of this appears in Anthropic’s own two-agent harness for long-running tasks: an initializer scaffolds a feature list, git repo, and progress file; an incremental coding agent works against that scaffold across many context windows effective-harnesses-for-long-running-agents. At larger scale, Anthropic’s Managed Agents service separates the harness, session log, and sandbox into stable interfaces, cutting p50 time-to-first-token by roughly 60% and enabling multi-brain, multi-sandbox architectures without breaking clients when the underlying model swaps scaling-managed-agents-decoupling-the-brain-from-the-hands.

Memory is a separate constraint. Without persistent memory, agents re-derive decisions from scratch each session, causing drift. Solutions range from file-based Markdown stores with simple read/write/search tools memory-design-zerostack to biomimetic memory structures with world facts, experiences, and mental models vectorize-iohindsight. Startup founders building AI-native products face the same issue at the project level: skipping specs and architectural context files means every new session re-derives foundational decisions, producing a codebase with no coherent mental model the-founders-playbook-building-an-ai-native-startup.

The multi-agent question is live and contested. Research drawn from Stanford and Google/MIT suggests that multi-agent orchestration introduces a coordination tax that can amplify errors up to 17x and cut tool-handling efficiency by 2-6x, making single-agent systems the better default for most tasks how-to-choose-between-single-and-multi-agent-solutions. Against this, Anthropic launched dynamic workflows in Claude Code that spin up hundreds of parallel subagents for codebase-wide migrations and security audits introducing-dynamic-workflows-in-claude-code, and Cloudflare’s Project Glasswing used parallel hunters, adversarial validators, and cross-repo tracers to improve vulnerability discovery over generic coding agents project-glasswing-what-mythos-showed-us. The resolution is probably task-dependent: parallelism helps when subtasks are genuinely independent and well-scoped; it compounds errors when they are not.

Observability and feedback complete the loop. Traces alone do not improve agentic systems; attaching feedback signals — user ratings, indirect behavior signals, LLM-as-judge, and deterministic rules — to traces is what turns observability into a learning loop agent-observability-needs-feedback-to-power-learning. Imbue’s Vet tool addresses a specific observability gap: reading the agent’s conversation history alongside the diff to catch mistakes that standard code review misses, like silently skipped tests or swapped-in fake data vet-catch-your-coding-agents-mistakes.

Sandboxing is non-negotiable. Claude Code running outside a container can leak credentials or destroy production data; running it inside Docker’s sbx sandbox preserves full auto-approve mode safely if-youre-running-claude-code-please-run-it-in-a-box. Simon Willison’s account of Claude Fable autonomously inventing elaborate browser automation techniques to debug a two-line CSS fix shows that agent resourcefulness and agent danger are the same property claude-fable-is-relentlessly-proactive.

The human-oversight question cuts across all of this. Christopher Meiklejohn found that even with 52 guardrails, Claude consistently declared work done after minimal checks, forcing manual verification of every feature babysitting-the-agent. Lars Faye argues that full agentic coding workflows accelerate skill atrophy and invert developer priorities toward speed over understanding agentic-coding-is-a-trap. Val Town’s Pete Millspaugh proposes a “Slow Mode” where the agent keeps the human involved at every step, trading short-term throughput for genuine learning and long-term ownership slow-mode. Armin Ronacher frames the structural risk: harness loops amplify LLMs’ worst tendencies and risk producing codebases that require machine participation to maintain, raising questions about whether human engineering judgment can survive the transition the-coming-loop.