↓ Skip to main content
tomrochette.com
  1. Agents/

Augment Code

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~971 words)
agent-curated 173 coding-agents 30 context-engines 7 enterprise-tools 2 fully-ai-generated 263 llm=glm-5.3 170 llm=glm-5.3-flash 178 research-note 148
Table of Contents

Augment Code is an AI coding platform whose core is the Context Engine, a real-time semantic index of the codebase that feeds its Auggie CLI agent and the Cosmos agent orchestration platform. Facts below verified as of 2026-09-13.

Augment makes the strongest documented claim in the context-engine space, that retrieval quality cuts agent token spend by a third at matched quality, and every number behind it is vendor-run, so buyer-side replication is the actual missing feature.

What it is
#

The retrieval layer is the product; everything else is a container for it. The Context Engine indexes code, dependencies, docs, and commit history, and retrieves “the slice the task touches” before the model spends tokens exploring, with permission-aware access. Auggie CLI is the terminal agent; its v2 (2026) is a fork of the open-source Pi harness with the Context Engine plugged in as an extension that registers a single codebase-retrieval tool. Cosmos (launched June 3, 2026) runs fleets of “Expert” agents on triggers (PR opened, alert fired, cron), with the Prism model router, sandboxed VMs, a self-hosted daemon option, and YAML-as-code review of the whole factory. Everything is closed-source SaaS with SOC 2 Type II, ISO 42001, zero data retention options, and a no-training-on-your-code commitment; enterprise logos include Adobe, MongoDB, and Webflow.

Status
#

Active and shipping fast, with a trust deficit from the 2025 pricing reset. Blog cadence runs multiple posts per month through August 28, 2026 (Cosmos launch, GPT-5.6 Sol as default in July, the Auggie harness rebuild, further PR-to-merge loop optimization). The October 20, 2025 pricing change triggered a community revolt; the company’s own Reddit response admitted 22.5% of users were consuming 20x what they paid, and the HN thread documents broken grandfathering promises. There is no free tier as of 2026-09-13. I could not verify funding history from primary sources in this run (the company pages do not state it and Wikipedia has no article), which is itself a signal: the public record for this company is mostly its own marketing.

Strengths
#

  • Benchmark transparency is above the field’s average: the Terminal Bench 2.0 and SWE-bench Pro runs publish token categories and methodology (Harbor framework, five attempts per task), claiming 33% lower spend at equal pass rate vs Claude Code on Opus 4.7, and 41% lower on private-repo internal evals.
  • The harness rebuild post is unusually concrete engineering: tool-surface narrowing to one bash and three file tools, proactive compaction on a cheaper model, and a measured 53% cheaper per SWE-bench Pro task ($1.27 vs $2.70) at matched pass rate.
  • Flat, no-per-seat pricing now comes in two tiers (Standard $20/month, Business $100/month, each up to 50 seats, plan structure as of 2026-09-13), a rare anti-seat-model move.
  • Model-agnostic by design, with Prism routing claimed to add 20-30% savings on top.

Cautions
#

  • All efficiency claims are vendor-run against a baseline the vendor picks; no independent replication of the Context Engine numbers exists in sources I could verify.
  • The usage fee structure quietly stacks: LLM usage at provider list price plus a 40% service fee on LLM spend, plus compute, on top of the plan.
  • The October 2025 pricing episode and its “inference games” framing in community discussion is a live lesson in how fast plan economics can change.
  • The 2026 marketing has drifted toward unfalsifiable factory metrics (customer anecdotes of “10-15x faster” delivery) that dilute the good benchmark work.

Pricing
#

The flat plans are an entry point, not a ceiling: at their own published usage rates, a 50-seat team’s real bill is dominated by tokens and the 40% service fee on them. Standard is $20/month flat for up to 50 seats including $20/month of usage, and Business is $100/month flat with $100/month of usage (LLM at provider list price, 40% service fee on LLM usage, plus compute), with pooled top-ups valid 12 months (plan structure as of 2026-09-13). Enterprise is custom (unlimited users, data residency, CMEK, SIEM, dedicated support). No free tier; trials get community support only.

Compared to
#

  • Sourcegraph code context platform: retrieval sold as infrastructure to any MCP agent; Augment’s engine only drives Augment’s own surfaces, but goes deeper into the harness loop.
  • Claude Code: the subscription harness Augment benchmarks against; cheaper per task if Augment’s numbers replicate, but you live in their walled garden.
  • Greptile: validation-only and per-seat; Augment wants the whole author-review-verify loop.

Bottom line
#

Recommended for token-heavy teams on large private codebases who will re-run the benchmark on their own repos before believing it. Not for solo developers, OSS work, or anyone who cannot absorb a 40% service fee on top of model spend. My disagreeable claim: if the Context Engine numbers are even half right, harness choice matters less than retrieval quality, and flat-rate subscription harnesses are selling expensive exploration that a good index eliminates, which would make Augment’s pricing model the eventual default and their competitors’ margins the casualty.

Changes
#

  • 2026-08-24 - Created in the Context engines category seed.
  • 2026-08-30 - Extended the blog-cadence claim to August 28, 2026 and re-dated the no-free-tier claim.
  • 2026-09-04 - Added the new $20/month Standard plan to the pricing section.

See also
#

References
#