Context Management in Coding Agents

The finding in 60 seconds

1. The context window is an attention budget, not a bucket.

   Chroma tested 18 models across eight lengths and found performance generally degraded as input grew, with non-uniform, model-specific curves S. Correct but irrelevant tokens can still reduce performance.

2. Subtraction beats addition, and the evidence is lopsided.

   5K tokens of targeted retrieval beat a 100K summary P. Masking matched summarisation at lower cost S. Fewer tools beat more tools S. A first audit should test for waste before changing strategy.

3. Compression damage shows up as unreliability before it shows up as lower accuracy.

   Compressed agents become intermittent: they solve a task, then fail the same task on a rerun S. Teams measuring single runs conclude compaction is free until production disagrees.

4. The cache can flip the sign of an optimisation.

   In the worked price model, a 24% token cut that made the prefix dynamic raised cost 6.9× C. Token count alone is not a cost model.

5. There is a stopping point.

   About 16 hours of sequenced work, most of it deletion, covers the large majority of the benefit. Past the stop rule, your binding constraint is somewhere else.

Choose your path

Pick the entry point that matches the time you have. Every path ends with something you can do.

The learning path

Five parts, one arc: understand the problem, learn the methods, measure them, make the hard calls, then act. Each chapter answers one question.

1. Part I

   Understand the problem

   What context is mechanically, when more of it makes agents less reliable, and where the tokens actually go.

   2 chapters · 16 min

   1. 1Why more context can hurt coding agentsWhat is context, mechanically, and when does adding more make coding agents less reliable?10 min
   2. 2Anatomy of a coding agent's contextWhere do the tokens in a coding agent's context actually go, and which parts can you control?6 min
2. Part II

   The methods

   The ten methods that survive scrutiny, how to choose among them, and the four surfaces where context is spent.

   6 chapters · 68 min

   1. 3Ten methods that survive scrutinyWhich context-management methods survive scrutiny, why does each work, and what does each cost?23 min
   2. 4Choosing your methods: a 55-minute procedureGiven my own agent and workload, which method should I adopt first, and in what order after that?5 min
   3. 5Retrieval: how code should enter contextShould code enter context by grep, embeddings, the symbol graph or agentic search, and what does the evidence actually say?9 min
   4. 6Compaction, offload and memoryHow do compaction, offload and memory lose information, what does that cost in measured terms, and what should you do with your wiki?15 min
   5. 7Sub-agents and isolationWhen does splitting work across sub-agents help, when does it destroy the task, and how can you tell in advance?8 min
   6. 8The tool surface: MCP, skills and progressive disclosureHow much context do tools, MCP servers and skills consume, and what does a healthy tool surface look like?8 min
3. Part III

   Measure it

   How to prove a context change worked, what to track, and what context really costs once the cache is priced in.

   2 chapters · 17 min

   1. 9Evaluating a context changeHow do you prove a context-management change helped, with what statistical design, and at what cost?9 min
   2. 10Metrics and the real cost of contextWhat should you measure, and what does context really cost once the cache, latency and people are counted?8 min
4. Part IV

   Decide and avoid

   The hard calls with no clean answer, the ways context management fails, the traps that look like hygiene, and real cases.

   4 chapters · 38 min

   1. 11Twenty-five hard callsWhich context decisions have no clean answer, and how do you decide them for your own situation?12 min
   2. 12Forty-four failure modesHow does context management fail, what is the tell for each failure, and how do you contain it?8 min
   3. 13Sixteen antipatternsWhat looks like good context hygiene but is not, and how do you tell the difference?8 min
   4. 14Eight case studiesWhat do these decisions look like on real, documented incidents, including ones where the change made things worse?10 min
5. Part V

   Act on it

   A sequenced 16-hour plan, a scored diagnostic, daily habits and the tooling landscape.

   4 chapters · 27 min

   1. 15Your optimisation planGiven everything so far, what exactly should I do, in what order, and when should I stop?9 min
   2. 16The diagnostic: score yourself, get one actionWhere does my team stand on context management today, and what is the single next thing to do?4 min
   3. 17Sixty-two habits that compoundWhat small, daily practices pay for themselves, and when does each one apply?6 min
   4. 18The tooling landscapeWhich open-source tools, plugins and local MCP servers improve context use, and which are worth adopting?8 min
6. Part A–B

   Reference

   Copyable templates and the full list of falsifiable lessons.

   2 chapters · 13 min

   1. ATemplates you can copyWhat are the fill-in forms that make this operational, and which field in each carries the value?3 min
   2. BFifty-eight lessons, each falsifiableWhat are the transferable lessons, why is each true, and what would prove each one wrong?10 min

Reference