Building Coding Agent Systems

The second season of Amp's documentary podcast follows the team's move from locally managed coding agents to remote-controllable, parallel agents running in isolated environments. The first three episodes cover Amp Neo, ephemeral orbs, exhaustive verification, production feedback loops, 'jellyware,' and how cheaper intelligence could change software economics, build-versus-buy decisions, and the work developers own.

Adam Azzam on dev boxes, resumable sessions, secret isolation, and control-plane design

Modal’s Adam Azzam argues that the hard part of background coding agents is the environment around the harness. He distinguishes bare sandboxes from task-ready dev boxes, explains why agent sessions need persistence and fast resumption rather than CI-style reruns, and recommends versioned per-repository images, secret-injecting proxies, and a separate control plane. The Ramp scale, pull-request share, and startup-time figures are first-party reports in a Modal product talk.

Shared agent infrastructure, pre-CI validation, and managed maintenance loops

Uber engineers Uday Kiran Medisetty and Adam Huda present the shared building blocks of a managed software factory: model and MCP gateways, pre-provisioned DevPods, a skills marketplace, a context graph, and the Cortana assistant. A staged walkthrough runs from product exploration through agent implementation, self-healing CI, and evidence-rich review. The adoption, token-saving, and quality figures are first-party reports without baselines or an independent security assessment; the speakers say the components are at different stages of rollout.

0xSero on Pi's provider abstraction, agent loop, tools, extensions, and local-model workflows

0xSero gives a live source-level tour of Pi: provider normalization, the streamed agent loop, built-in tools, extensions, session storage, and the SDK. He then shows how those pieces can support local models, delegated agents, and long-running goal loops. The cache-hit, latency, cost, and benchmark comparisons are demonstrations without enough linked configuration or methodology to count as controlled evaluations.

Mario Zechner and Armin Ronacher on Pi's small core, self-extension, and open ecosystem

CultRepo profiles Mario Zechner’s path from open-source game tooling to Pi, a deliberately small coding-agent harness. Zechner explains Pi’s nested tool loop, four-package TypeScript architecture, and self-extension model; he and Armin Ronacher connect minimalism to predictable tools, human responsibility, and model neutrality. The film’s adoption figures and leaderboard placement are first-party reports without defined scope, and its open-foundation plans describe an ambition.

Continuous AI, bounded repository automation, cost controls, and evidence-rich review

GitHub Principal Researcher Don Syme defines continuous AI as recurring, shared agent work that runs beside deterministic CI/CD. He explains GitHub Agentic Workflows: bounded repository context, permission and security controls, scheduled cost management, concurrent model evaluation, and quality gates that give human reviewers evidence. The claims are first-party reports about a public-preview product, without an independent security assessment or comparative data.

Continual learning, the Big World Hypothesis, and a research path beyond frozen model weights

Rich Sutton and Khurram Javed argue that deployed LLMs stay static because their weights do not update from experience. They propose continual learning from the real world, per-weight step-size optimization, Continual Backprop, and self-discovered abstractions as a path toward agents that keep adapting. Their claims about synthetic data, forgetting, and replacing the current LLM paradigm describe Oak Lab’s research agenda, not a system demonstrated at frontier scale.

Context-budgeted plugins, event-driven feedback, and isolated coordination across persistent agents

Claude Code engineer Daisy Hollman explains a context-budgeted approach to agentic engineering beyond one repository: audit external work systems, treat context and prefix-cache costs as architecture constraints, and prefer lazy or event-driven context. She compares MCP tools, skills, subagents, and hooks, then shows her worktree, messaging, and scheduled-wake-up setup. The cache, cost, and throughput claims are presenter or vendor reports without comparative methods; several roadmap details are tentative.

Anthropic engineers on adaptable harnesses, durable sessions, isolated execution, and production agent infrastructure

Anthropic Applied AI engineers Gagan Bhat and Isabella Kai He explain why agent harness assumptions expire as models change, how decoupling the model loop from tool execution improves recovery and isolation, and how durable session logs support context management and observability. The talk is an Anthropic product presentation, so its latency measurements and proposed memory and grading features remain vendor-reported.

An Italian-language research notebook on model representations, reasoning, interpretability, architecture claims, open weights, local inference, efficiency, and speculative questions about AI and cosmology. The guided summary frames the playlist as conceptual grounding for evaluating coding agents rather than as a settled theory of AI.

A selected coding-agent route through Mayank Gupta's broader interview playlist, covering agent harnesses, review, architecture, model economics, context, and the changing work of software developers. The guests remain the sources for their individual claims.

Boris Cherny on prompt ablation, model elicitation, and long-running agent workflows

Boris Cherny explains why Anthropic removed 80% of Claude Code's system prompt for Opus 5, how to rebuild instructions through empirical ablation, why hard tasks need strong verification, and how dynamic workflows and recurring routines orchestrate agents for long-running work.

A portable concept-and-bundle contract for agent knowledge, with reference producer and consumer tools

Prism Labs explains Google Cloud’s v0.1 Open Knowledge Format proposal: concepts are Markdown files with minimal YAML metadata, ordinary links between them form a graph, and optional index and log files add navigation and history. Google supplies a BigQuery enrichment agent, a graph visualizer, sample bundles, and Knowledge Catalog ingestion as first-party proofs of concept. The recording shows no independent cross-vendor interchange or retrieval-quality evidence, and it misstates which fields are required.

Mario Zechner on building and using Pi

Mario Zechner demonstrates how he uses Pi to maintain Pi itself, pairing a deliberately minimal agent harness with strong human architectural control. The session covers minimalist tool design, self-modifying agent extensions, issue triage with GitHub CLI, deterministic checks, transcript-like branch workflows, and why human taste and module boundaries still matter when agents can generate code at high volume.

Mario Zechner on Pi, self-modifying agents, and agent slop

Mario Zechner explains why he built Pi as a minimal, self-modifying coding-agent harness after frustrations with opaque context management and feature-heavy tooling. The talk covers Pi's small core and extension model, Terminal-Bench results, agent spam pressure on open-source maintainers, and a practical warning to slow down, scope agent tasks, and keep humans responsible for critical code.

Mario Zechner on why Pi strips coding agents down to a malleable core

Mario Zechner explains why he built Pi after Claude Code, OpenCode, and other harnesses stopped matching his need for predictable context, observability, and extensibility. The talk covers agent history from ChatGPT copy-paste to agentic search, Claude Code feature bloat, OpenCode context pitfalls, Terminal-Bench and Terminus as evidence for minimal harnesses, Pi’s four-tool design and TypeScript extensions, and defenses against agent-generated open-source spam.

A multi-episode series in which the Amp team documents building Amp. Topics include the shift from prompt to context engineering, autonomous feedback loops, sub-agent architectures for context management, the ‘Oracle’ pattern for multi-model reasoning, why codebases must adapt to agents, and evaluating models across providers. The team argues that results depend on model quality rather than hidden tricks, and treats agents as power tools rather than magic.